WO2017218492A1 - Neural decoding of attentional selection in multi-speaker environments - Google Patents
Neural decoding of attentional selection in multi-speaker environments Download PDFInfo
- Publication number
- WO2017218492A1 WO2017218492A1 PCT/US2017/037186 US2017037186W WO2017218492A1 WO 2017218492 A1 WO2017218492 A1 WO 2017218492A1 US 2017037186 W US2017037186 W US 2017037186W WO 2017218492 A1 WO2017218492 A1 WO 2017218492A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- signals
- neural
- speaker
- person
- sound sources
- Prior art date
Links
- 230000001537 neural effect Effects 0.000 title claims abstract description 94
- 238000000034 method Methods 0.000 claims abstract description 94
- 230000005236 sound signal Effects 0.000 claims abstract description 86
- 238000000926 separation method Methods 0.000 claims abstract description 81
- 238000012545 processing Methods 0.000 claims abstract description 60
- 238000013528 artificial neural network Methods 0.000 claims abstract description 42
- 230000008569 process Effects 0.000 claims description 22
- 238000010219 correlation analysis Methods 0.000 claims description 13
- 238000002566 electrocorticography Methods 0.000 claims description 7
- 238000005259 measurement Methods 0.000 claims description 5
- 210000004761 scalp Anatomy 0.000 claims description 5
- 230000001131 transforming effect Effects 0.000 claims description 3
- 239000000203 mixture Substances 0.000 description 75
- 238000013459 approach Methods 0.000 description 34
- 238000002474 experimental method Methods 0.000 description 23
- 238000012549 training Methods 0.000 description 23
- 230000006870 function Effects 0.000 description 17
- 238000004458 analytical method Methods 0.000 description 12
- 239000010410 layer Substances 0.000 description 9
- 238000012360 testing method Methods 0.000 description 9
- 230000003416 augmentation Effects 0.000 description 7
- 230000008901 benefit Effects 0.000 description 7
- 238000011156 evaluation Methods 0.000 description 7
- 230000015654 memory Effects 0.000 description 7
- 238000010606 normalization Methods 0.000 description 6
- 210000004556 brain Anatomy 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 230000002452 interceptive effect Effects 0.000 description 5
- 230000002123 temporal effect Effects 0.000 description 5
- 238000004590 computer program Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 230000002093 peripheral effect Effects 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000007704 transition Effects 0.000 description 4
- 238000010200 validation analysis Methods 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 210000003484 anatomy Anatomy 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 238000001793 Wilcoxon signed-rank test Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008904 neural response Effects 0.000 description 2
- 230000000926 neurological effect Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000006403 short-term memory Effects 0.000 description 2
- 230000003595 spectral effect Effects 0.000 description 2
- 208000023914 Central Auditory disease Diseases 0.000 description 1
- 241001212789 Dynamis Species 0.000 description 1
- 208000032041 Hearing impaired Diseases 0.000 description 1
- XDXHAEQXIBQUEZ-UHFFFAOYSA-N Ropinirole hydrochloride Chemical compound Cl.CCCN(CCC)CCC1=CC=CC2=C1CC(=O)N2 XDXHAEQXIBQUEZ-UHFFFAOYSA-N 0.000 description 1
- SEQDDYPDSLOBDC-UHFFFAOYSA-N Temazepam Chemical compound N=1C(O)C(=O)N(C)C2=CC=C(Cl)C=C2C=1C1=CC=CC=C1 SEQDDYPDSLOBDC-UHFFFAOYSA-N 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 230000002238 attenuated effect Effects 0.000 description 1
- 210000003926 auditory cortex Anatomy 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 239000004020 conductor Substances 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007917 intracranial administration Methods 0.000 description 1
- 230000002045 lasting effect Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000013518 transcription Methods 0.000 description 1
- 230000035897 transcription Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000011870 unpaired t-test Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0272—Voice signal separating
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/01—Input arrangements or combined input and output arrangements for interaction between user and computer
- G06F3/011—Arrangements for interaction with the human body, e.g. for user immersion in virtual reality
- G06F3/015—Input arrangements based on nervous system activity detection, e.g. brain waves [EEG] detection, electromyograms [EMG] detection, electrodermal response detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/16—Sound input; Sound output
- G06F3/165—Management of the audio stream, e.g. setting of volume, audio stream path
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; DEAF-AID SETS; PUBLIC ADDRESS SYSTEMS
- H04R25/00—Deaf-aid sets, i.e. electro-acoustic or electro-mechanical hearing aids; Electric tinnitus maskers providing an auditory perception
- H04R25/50—Customised settings for obtaining desired overall acoustical characteristics
- H04R25/505—Customised settings for obtaining desired overall acoustical characteristics using digital signal processing
- H04R25/507—Customised settings for obtaining desired overall acoustical characteristics using digital signal processing implemented by neural network or fuzzy logic
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L2021/02087—Noise filtering the noise being separate speech, e.g. cocktail party
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; DEAF-AID SETS; PUBLIC ADDRESS SYSTEMS
- H04R3/00—Circuits for transducers, loudspeakers or microphones
Definitions
- a method includes obtaining, by a device, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources.
- the method further includes obtaining, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, selecting one of the plurality of separated signals based on the obtained neural signals for the person, and processing the selected one of the plurality of separated signals.
- Embodiments of the method may include at least some of the features described in the present disclosure, including one or more of the following features,
- the method may further include generating a sound spectrogram from the combined sound signal, and applying the neural-network-based speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
- Selecting the one of the plurality of separated sound signals based on the obtained neural signals for the person may include generating an attended speaker spectrogram based on the neural signals for the person, comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms, and transforming the selected one of the multiple resultant speaker spectrograms into an acoustic signal.
- Comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms may include comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
- Obtaining the neural signals for the person may include obtaining one or more of, for example, electrocorticography (ECoG) signals for the person, neural measurements via a noninvasive scalp, and/or in-ear EEG recordings.
- ECG electrocorticography
- Applying the neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the combined sound signal.
- DNN deep neural network
- Processing the selected one of the plurality of separated sound signals may include one or more of, for example, amplifying the selected one of the plurality of separated signal s, and/or attenuating at least one non-selected signal from the plurality of separated signals.
- Obtaining the combined sound signal for the multiple sound sources may include receiving the combined sound signal for the multiple sound sources at a single microphone coupled to the device.
- the combined sound signal may define a time-frequency embedding space.
- Applying the neural-network-based speech-separation processing to the combined sound signal may include determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sound sources in the embedding space for the combined sound signal, deriving masks for the determined reference points, and extracting at least one of the multiple sound sources using at least one of the derived masks.
- Determining the respective reference points may include determining the respective reference points using a deep neural network.
- Deriving the masks may include computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points.
- the reference points may include attractor points in the embedding space.
- a system in some variations, includes at least one microphone to obtain a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, one or more neural sensors to obtain neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, and a controller coupled to the at least one microphone and the one or more neural sensors.
- the controller is configured to apply neural -network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
- non-transitory computer readable media are provided, thai are programmed with instructions, executable on a processor, to obtain, by a device comprising the processor, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and apply, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources.
- the instructions further include one or more instruction, executable on the processor, to further obtain, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
- an apparatus includes means for obtaining a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and means for applying neural -network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources.
- the apparatus further includes means for obtaining neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, means for selecting one of the plurality of separated signals based on the obtained neural signals for the person, and means for processing the selected one of the plurality of separated signals.
- Embodiments of the system, the computer-readable media, and the apparatus may include at least some of the features described in the present disclosure, including at least some of the features described above in relation to the method.
- FIG. 1 is a diagram of a system to process (e.g., separate) combined sound signals (e.g., for attend onal selection applications).
- FIG. 2 is another example implementation of a system to process (e.g., separate) sound signals,
- FIG. 3 is a flowchart of an example procedure for attentional selection of a speaker in a multi-speaker environm ent.
- FIG. 4 is a graph providing an analysis of deep neural network (DNN) output correlation.
- DNN deep neural network
- FIG. 5 includes graphs illustrating reconstruction accuracy for neural correlation analysis.
- FIG. 6A includes graphs showing the proportion of segments in which the attentional focus of subjects could be correctly determined using HF data and LF data.
- FIG. 6B includes graphs presenting attention decoding index (ADI) measures per electrode.
- ADI attention decoding index
- FIG. 7 is a graph showing results of an analysis for dynamic switching of attention for an example subject.
- FIGS. 8A-D are graphs showing further performance results in relation to dynamic switching of attention
- FIGS, 9A-B are graphs illustrating speaker separation quality results.
- FIG. 10 is a diagram of a proposed system architecture to implement a deep-attractor network (DANet) approach.
- DANet deep-attractor network
- FIG. 11 A includes graphs showing an example of a mixture, the difference between two speakers in the mixture, and the separated spectrograms of the two speakers using a DANet implementation.
- FIG. 1 IB is a graph showing locations of T-F bins in an embedding space.
- FIG. 12 is a graph presenting the locations of attractor points in the embedding space.
- FIG. 13 is a flowchart of an example deep attractor signal separation procedure.
- FIG. 14 is a schematic diagram of a generic controller system. [0039] Like reference symbols in the various drawings indicate like elements. DETAILED DESCRIPTION
- Described herein are systems, devices, apparatus, methods, computer program products, media, and other implementations, to process sound signals / data in environments that contain multiple sounds sources, based on neural signals / data from a person receiving the multi-source sound signals.
- procedures to decode which speaker a person is attending to are implemented by monitoring their neural activity via both invasive and non-invasive electrophysiological recordings.
- attention-aware brain-computer interfaces may be used to control smart bearable devices capable of selectively processing (e.g., amplifying) one speaker and/or suppressing other speakers in crowded environments.
- DNN refers to neural networks that include multiple hidden layers between the input and output layers of the networks.
- the output of the speech separation processes can be used to decode the attentional state of the listener, and subsequently, to amplify the attended source.
- a speech separation procedure separates out the speakers in a crowd, and compares to a signal reconstructed from the brain.
- the speech separation procedure can next amplify (or otherwise filter) that speaker source compared to other sources.
- implementations described herein include a method that comprises obtaining, by a device (e.g., a hearing device, a hearing enhancement device to allow better hearing in a very noisy environment, a noise cancellation device, a virtual reality device, and/or other types of hearing augmentation devices), a combined sound signal for sound combined from multiple sound sources in an area in which a person is located.
- the method further includes applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, and obtaining, by the device, neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to.
- the method additionally includes selecting one of the plurality of separated signals based on the obtained neural signals for the person, and processing the selected one of the plurality of separated signals.
- obtaining the neural signals for the person may include obtaining one or more of, for example, electrocorticography (ECoG) signals for the person, neural measurements via a non-invasive scalp, and/or in-ear EEG recordings.
- Applying the speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the multiple sound sources based on the combined sound signal.
- Processing the selected one of the plurality of separated sound signals may include one or more of, for example, amplifying the selected one of the plurality of separated signals, and/or attenuating at least one non-selected signal from the separated sound signals.
- separation processes to separate the multi-source audio / sound signals may be realized as a deep learning framework for single channel speech separation by creating attractor points in high dimensional embedding space of the acoustic signals, which pull together the time-frequency bins corresponding to each source. Attractor points are created, in some embodiments, by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings.
- the combined signal may be a mixture speech signal defining a time-frequency embedding space.
- Applying the speech- separation processing to the combined sound signal may include, in such embodiments, determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sources in the embedding space for the mixture speech signal, deriving masks for the determined reference points, and extracting at least one of the multiple sources using at least one of the derived masks.
- the system 100 (which may be implemented as a hearing device to facilitate hearing for a person 104, or may be implemented as a noise cancellation device, or any other type of a hearing augmentation device) includes a microphone 110 to receive a combined sound signal 102 combining sound signals from multiple sound sources.
- a controller e.g., a computing-based device, which may be incorporated into the hearing device, or some other hearing augmentation device, used by the person 104) implements one or more sound-processing procedures 120 (which may each correspond to a separate DNN, although, in some embodiments, a single neural network unit may be configured to implement multiple separate DNN's) to process the combined (aggregated) sound signal 102 to separate the combined sound signal into a plurality of separate signals 122a-n (which may include sound signals, or may include signal representative and/or derived from sound signals).
- the one or more procedures 120 implemented by the system's controller may include a speech-separation process based on a deep-neural network.
- the system 100 also includes one or more sensors to obtain / measure neural signals / data 106 from the person 104 that are indicative of the sound source (e.g., one of multipl e speakers in the vicinity of the person 104) the person 104 is attentive to. Based on the neural signals / data obtained for the person 104, one or more of the plurality of separated signals derived by the one or more procedures 120 is selected. In the example of FIG. 1, the neural signals obtained for the person 104 may indicate that the user is attentive to the sound source SP2, and, therefore, separated sound signal 122b (corresponding to the source SP2 in FIG. 1) is selected for further processing. The selected separated sound signal can then be processed to, for example, amplify, or otherwise perform signal enhancing filtering on the signal. Additionally and/or alternatively, the non-selected signals (i.e., the signals other than the sound signal 122b) may be processed to attenuate or suppress them (or to perform some other filtering operation thereon),
- FIG. 2 depicts another example implementation of a system 200 to process (e.g., separate) sound signals (e.g., in a closed-loop).
- two speakers namely, Spkl and Spk2
- the first stage 210 includes various signal recording and pre-processing operations performed on acquired audio signals.
- some source speakers were recorded using Tucker Davis Technologies (TDT ⁇ ) hardware and sampled at 2441Hz, while other source speakers were recorded using Xltek® hardware and sampled at 500Hz. Further processing operations were performed offline.
- filters were realized using Matlah's® Filter Design Toolbox, and were used in both a forwards and backwards direction to remove phase distortion.
- TDT acquired data
- the data was resampled to and a 1 st order Butterworth highpass filter with a cut-off frequency at IHz was used to remove DC drift.
- the data was subsequently re-referenced using a common average scheme, with the effect of noisy electrodes mitigated by using a trimmed mean approach (e.g., removing the top and bottom 10% of electrodes at each sample point).
- These example implementations also included removing line noise at 60Hz and its harmonics (up to 240Hz) using 2 nd order IIR notch filters with a bandwidth of IHz. A period of silence was recorded before each experiment, and the corresponding data was normalized by subtracting the mean and dividing by the standard deviation of the pre-stimulus period. These implementations also included filtering data into two frequency bands known to be responsive to speech, namely, the high-frequency (HF) gamma band (70-150Hz), and the low-frequency (LF) delta and theta bands (l-8Hz).
- HF high-frequency
- LF low-frequency
- the power of the high-frequency band and the phase of the low-frequency band are modulated by speech.
- the data was first filtered into 8 frequency bands between 70 and 150Hz, using, for example, Chebyshev Type 2 filters, each with a bandwidth of 10Hz.
- the power analytic amplitude of each band could then be derived using a Hilbert transform.
- the average of all 8 frequency bands can then be used to represent the total HF power.
- an 8 th order (4 sections) Butterworth low- pass filter was applied to the data with a corner frequency of, for example, 8Hz, and 20dB attention at 12Hz,
- signal separation processing is applied to the combined (mixed) audio signal (which may have been processed based on, for example, processing operation such as those described in relation to stage 210) at stage 220.
- speech separation processing is based on a deep neural network (DNN) approach.
- DNN deep neural network
- a spectrogram of the mixture is obtained.
- This spectrogram is then fed to each of several DNNs, each trained to separate a specific speaker from a mixture (DNN Spkl to DNN SpkN).
- DNN Spkl to DNN SpkN
- a user may be attending to one of the speakers (in this case, Spkl).
- a spectrogram of this speaker is reconstructed from the neural recordings of the user.
- This reconstruction is then compared with the outputs of each of the DNNs using a normalized correlation analysis (as shown at stage 230) in order to select the appropriate spectrogram, which is then converted into an acoustic waveform and added to the mixture so as to amplify the attended speaker (and/or attenuate the signals corresponding to the non-attended speakers).
- a normalized correlation analysis as shown at stage 230
- a single-channel speech separation that utilizes, in some embodiments, a class of DNNs known as long short-term memory (LSTM) DNNs is employed.
- LSTM long short-term memory
- Each DNN is trained to separate one specific speaker from arbitrary mixtures.
- two speakers Spkl F and Spk2M
- a system configured to work in a real -world situation would generally contain multiple DNNs, each trained to separate specific speakers, any of whom may or may not be present in the environment.
- 4 DNNs were trained to separate 4 speakers (two female speakers and two male speakers), hereafter referred to as SpklF, Spk2M, Spk3F, and Spk4M.
- ail speakers were native American English speakers.
- the speech waveforms were converted, in the example experiment, into 100-dimensional Mel-frequency spectrograms.
- the goal was then to obtain an estimate S' of a clean target spectrogram S from a mixture M.
- a soft mask Y' is learned and applied to the mixture to mask the interfering speech.
- the squared Euclidian distance between the masked spectrogram and the clean target spectrogram is treated as the error in order to generate the gradient that is back propagated through the DNN to update the parameters.
- An objective function that may be used is:
- the input to the DNNs was the logarithm (base 10) of the spectrograms, normalized so that each frequency band had zero mean and unit variance.
- Each DNN in the implementations realized and tested, contained four (4) layers with 300 nodes each, followed by a single layer containing 100 nodes with logistic activation to output a spectrogram.
- An acoustic waveform was generated by combining the spectrogram that was output by the DNN with the phase of the original mixture (e.g., inverting, in some embodiments, the spectrogram with the phase of the original mixture).
- each of the DNNs did not receive training sample data for the other target speakers during training. That is, in the implementations realized and evaluated herein, a DNN trained to separate one speaker (e.g., SpklF) was not exposed to training data for other speakers,
- DANet deep attractor network
- the a deep learning framework for single channel speech separation is used that is configured to create attractor points in high dimensional embedding space of the acoustic signals, which pull together the time-frequency bins corresponding to each source. Attractor points maybe created by finding the centroids of the sources in the embedding space, which are
- DANet DANet
- K-means strategy K-means strategy
- fixed-attractor- points strategy where the latter requires no post-processing and can be implemented in real-time.
- the system was configured to output several waveforms, each corresponding to one sound source in the mixture, in response to an input mixed waveform
- the processing operations of such a system implementation may include acquiring mixture speech signal with a (single) sensor, processing the signal to estimate a set of features, processing the features by a deep neural network (DNN) to assign a descriptor for each element in the signal, estimating the attribution of each element according to the physical properties of the descriptors, and recovering the target sources according to the estimation.
- DNN deep neural network
- selection of the speaker is performed through use of a speaker selection stage 230 and a stimulus reconstruction stage 240 of the system 200.
- the signals corresponding to the selected speaker are then processed (e.g., amplified) using the filtering processes implemented at stage 250, Additionally and/or alternatively, signals corresponding to non-selected speakers mat be processed to, for example, attenuate those signals.
- the stimulus-reconstruction procedure implemented by the stage 240 applies a spatiotemporal filter (decoder) to neural recordings to reconstruct an estimate of the spectrogram of the attended speaker,
- a 100-dimensional log Mel-frequency spectrograms maybe downsampled by a factor of 10, resulting in ten (10) frequency bands used for neural recordings.
- Each decoder can be trained using the data from the single- speaker (S-S)
- Electrodes may be chosen if they are significantly more responsive to speech than to silence. For the high gamma (HG) data, this means that the average power during speech is greater than during silence. For the low frequency (LF) data, electrodes may be chosen if the average power is either significantly greater or lower during speech than during silence.
- the data can be segmented into SOOms chunks and divided into two categories; speech and silence. Significance may then be determined using an unpaired t-test (false discovery rate (FDR) corrected, q ⁇ 0.05). This results in varying numbers of electrodes retained for each frequency band and each subject, The decoders may be trained using time-lags from -400 to 0 ms.
- Determining to whom the subject user is attending may be implemented through a correlation analysis, e.g., using Pearson's r- value (e.g., the linear correlation measure between two variables). Typically, whichever spectrogram has the largest correlation with the reconstructed spectrogram is considered to be the attended speaker. However, in embodiments in which four (4) DNNs are used, each trained to separate a speaker that may or may not be present in the combined signal (mixture), the analysis becomes slightly more complex. Because the DNNs that do not see their designated speakers in the mixture are likely to output a signal very close to the mixture, it may be necessary to normalize the correlation values with respect to the mixture. This is because the correlation between the reconstructed spectrograms and the mixture can be very large.
- the decoder configured to identify the specific speaker is referred to as 5DNNK
- the spectrogram of the mixture signal is represented as 5MIX
- the reconstructed spectrogram is applied to each
- DNN is normalized relative to the distribution of a scores from all DNNs, according to:
- the next neural correlation processing stage includes subtracting the correlation between and adding the constant 1, providing a resultant normalized correlation value,
- the reconstructed spectrogram is first segmented (from the multi-speaker experiments) into, for example, 20-second bins.
- the reconstructed spectrogram is first segmented (from the multi-speaker experiments) into, for example, 20-second bins.
- four normalized correlation values are produced by the DNN's for each segment (the normalized correlation values are denoted
- SpklF and Spk2M are the only speakers that are presented to the subject (the listener), it is expected that PI.
- ADI Attention Decoding Index
- the neural data was artificially divided and concatenated into consecutive segments in which subjects were attending to either speaker. Specifically, the data was divided into ten (10) segments, each lasting 60 seconds. The subjects attended to the male speaker for the first segment.
- a sliding window approach was used in which correlation values were obtained every second over a specified window. Window sizes ranging from 5 to 30 seconds (in 5 second increments for 6 window sizes in total) were used. Larger windows should lead to more consistent (i.e., less noisy) correlation values and provide a better estimate of the attended speaker. However, they should also be slower at detecting a switch in attention.
- implementations described herein include a system comprising at least one microphone to obtain a combined sound signal for sound combined from multiple sound sources in an area in which a person is located, one or more neural sensors to obtain neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, and a controller coupled to the at least one microphone and the one or more neural sensors.
- the controller is configured to apply speech-separation processing (e.g., neural-net processing such as deep neural network, or DNN) to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
- speech-separation processing e.g., neural-net processing such as deep neural network, or DNN
- the procedure 300 includes obtaining 310, by a device, a combined or mixed sound signal for signals combined from multiple sound sources in an area in which a person is located.
- the device of procedure 300 may include one or more of, for example, a hearing device, a hearing enhancement device (to allow better hearing in a very noisy environment), a noise cancellation device, a virtual reality device, and/or other types of hearing augmentation devices.
- the combined signal may be obtained through use of a single receiving sensor (e.g., a single microphone), although the procedure 300 may be performed using multiple receiving sensor which may be part of the device implementing the procedure 300.
- the procedure 300 includes applying 320, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources.
- Applying the neural-network- based speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the multiple sound sources based on the combined sound signal.
- DNN deep neural network
- the DNN may include a plurality of individual DNN's (e.g., long short-term memory (LSTM) DNN's) that are each configured to separate one respective specific speaker from an arbitrary mixture of signals. Configuring the individual DNN's to separate their respective specific speakers may be realized through an initial training of the DNN's using speech data from the specific speaker, and speech data from interfering speakers (e.g., a combined / mixed speech signals).
- speech separation processing may be applied to resultant data derived from the raw combined speech data (signals). For example, the speech separation processing may be applied to sound spectrogram data to derive multiple resultant speaker spectrograms.
- the procedure 300 may further include generating a sound spectrogram from the combined sound signal, and applying the speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
- the combined sound signal may define a time-frequency embedding space
- applying the speech-separation processing to the combined sound signal may include determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sources in the embedding space for the combined sound signal, deriving masks for the determined reference points, and extracting at least one of the multiple sound sources using at least one of the derived masks.
- determining the respective reference points may include determining the respective reference points using a deep neural network. Deriving such masks may include computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points.
- the reference points may include attractor points in the embedding space.
- the speech-separation processes described herein are discussed primarily in relation to attentional selection application (e.g., to enhance the speech of a speaker, from a plurality of speaker, that the listener is focusing on), the speech-separation processes described herein may be utilized for other application.
- the speech separation processes may ⁇ be used in surveillance application (to separate and record speech from a combined sound signal corresponding to different speakers).
- the speech separation processes described herein may be used to automatically analyze and transcribe survilance audio, which can be hundreds of hours long.
- the speech separation processes can be used as a preprocessing step, where the target speech sounds are first separated from other sources that might have been present during the recording, and the cleaned speech signal can be passed to speech recognition implementation.
- the speech separation processes described herein may be used for virtual reality applications, in which speech separation techniques discussed are used to reconstruct speaking objects from a pre-recorded source (and render the separated speech signals into a virtual audio (or audio-visual) scene).
- speech separation techniques discussed are used to reconstruct speaking objects from a pre-recorded source (and render the separated speech signals into a virtual audio (or audio-visual) scene).
- One of the challenges in 3D virtual audio is that rendering an acoustic scene requires separate audio tracks for individual sources, so that the sound sources can be processed separately and placed in different positions relative to the listener.
- the speech processing procedures described herein can be used to first separate all the sources in the recording, and subsequently re-syntehsize them with desired location and attribute.
- the procedure 300 also includes obtaining 330, by the device, neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to.
- Obtaining the neural signals for the person may include obtaining one or more of, for example, eiectrocorticography (ECoG) signals for the person, neural measurements via a non-invasive scalp, and/or in-ear EEG recordings.
- EoG eiectrocorticography
- the procedure 300 includes selecting 340 one of the plurality of separated signals based on the obtained neural signals for the person.
- selecting the one of the plurality of separated signals may include generating an attended speaker spectrogram based on the neural signals for the person, comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms, and transforming the selected one of the multiple resultant speaker spectrograms into an acoustic signal.
- comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms may include comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
- the procedure 300 further includes processing 350 the selected one of the plurality of separated signals.
- processing the selected one of the plurality of separated sound signals may include one or more of amplifying the selected one of the plurality of separated signals, and/or attenuating at least one non-selected signal from the separated signals.
- FIG. 4 includes a graph 400 providing an analy sis of the DNN output correlation.
- a graph 400 providing an analy sis of the DNN output correlation.
- the results from the four networks are split into two by averaging the r-values obtained from the networks that were trained to separate a femal e speakers (left porti on 410 of the graph 400 ) or male speakers (right portion 420 of the graph 400).
- the bars 412 and 422 show the results when a DNN was presented with a mixture containing its designated pre-trained speaker (trained target), while the bars 414, 416, 424, and 426 are the results when the mixtures contained an undesignated speaker (i.e., an untrained target, with the bars 414 and 424 corresponding to female speakers, and 416 and 426 corresponding to make speakers),
- the dotted line 430 shows the average correlation between the raw mixture and the clean target speaker.
- FIG. 5 includes graphs 500, 510, 520, and 530, illustrating reconstruction accuracy for the neural correlation analysis.
- the left part of FIG. 5, including the graphs 500 and 510, provides the reconstruction accuracy in terras of raw r-values, and shows the average correlations between the reconstructed spectrograms and the outputs of the DNNs for each subject and neural frequency band (where each subject is represented by a dot of a different shade).
- the r-values labeled as attended and unattended come from the DNNs trained on SpklF and Spk:2M
- the r- values labeled as undesignated come from the DNNs that were trained for Spk3F and Spk4M (as used herein, "undesignated” means that the DNN's in question were not trained to separate either speakers in the mixture that the listener actually listened to).
- the attended r-values are typically higher than the unattended r-values, there is also a high correlation between the reconstructed spectrograms and the mixture. Because the DNNs that do not see their designated speaker usually output spectrograms similar to the mixture, the undesignated correlation values were also relatively high.
- the r-values are normalized with respect to the mixture.
- An example normalization solution involves three key steps; (i) incorporating the mixture into the correlation analysis (in accordance with, for example, Equations 2 and 3 above), (ii) performing a test-normalization (t-norm) to equalize the outputs of each DNN with respect to each other (in accordance with, for example, Equation 4, above), and (iii) subtracting the correlation between the DNN output and the mixture (in accordance with, for example, Equation 5).
- the graphs 520 and 530 at the right portion of FIG. 5 provide the correlation analysis results in terms of normalized r-values, and indicate thai after applying the normalization processes, the attended correlation values were far higher than either the unattended or undesignated r-values, which allows adequate decoding of the attended speaker.
- FIG. 6A includes graphs 610 and 620 showing the proportion of segments (20s) in which the attentional focus of each subject could be correctly- determined using HF data (the graph 610 at the left part of FIG. 6 A) and LF data (the graph 620 at right part of FIG. 6A).
- the attentional focus of a subject was considered to be successfully decoded if the ADI was greater than the 0.45 threshold (indicated by the lines 612 and 622) calculated using a random shuffle of the data.
- the attentional focus of subjects 1, 2, and 3 could be decoded using HF (70- 150Hz) data, and subjects 1, 2 and 5 using LF (l-8Hz) data. Because different regions in the auditory cortex are differentially modulated by attention, the variability in ADI across subjects was thought to be related to the anatomical locations of electrodes.
- the pie charts in the graphs 610 and 620 illustrate the proportion of electrodes that were responsive to speech from two (2) anatomical regions: Heschl's gyrus, and the superior temporal gyrus. Electrodes that were responsive to speech, but that were not in either of these locations, are collectively referred to as Other. These Other electrodes were located in various anatomical regions including the middle and inferior temporal gyri, and planum temporale.
- the single number displayed above each pie chart refers to the total number of electrodes that were responsive to speech for that subject and for the corresponding frequency band.
- FIG. 6B which includes graphs 630 and 640 presenting ADI measures per electrode. Similar to FIG. 6A, electrodes are shaded according to anatomical location: Heschi's gyms (FIG; medium dark shade), superior temporal gyrus (STG; dark shade), and Other (light shade).
- FOG Heschi's gyms
- STG superior temporal gyrus
- Other light shade.
- HF Heschi's gyms
- STG superior temporal gyrus
- Using LF data no region produced ADIs significantly greater than zero (p > 0.05).
- FIG. 7 shows the results of the analysis for dynamic switching of attention for an example subject (subject 1) using HF data and a 20 second window size. Each solid vertical line marks a switch in attention.
- Portion 710a-e of the graph correspond to segments when the subject was attending to Spk2m, and portions 720a-e correspond to segments when the subject was attending to Spklf. Normalized correlation values for each of " the 4 DNNs are plotted at the bottom of the figure.
- FIG. 8A displays the same results as shown in FIG. 7, but averaged over all sections when the subject was attending to Spk2m (-60s - 0s) and Spklf (0s - 60s). Shaded regions denote standard error.
- This analysis was used in order to determine a measure of the transition time (how long it takes to detect a switch in attention for a subject). Transition times were calculated as the time at which the blue (Spk2M) and red (SpklF) lines intersected in the averaged data. This calculation was performed for each subject whose attentiona! focus could be decoded and for all window-sizes (between 5 and 30 seconds: see FIG. 8B). For all subjects, the transition-time increases monotonically with larger window-sizes.
- FIG. 8C shows an example of the same analysis as in FIG. 8A, but instead uses LF data from subject 2.
- FIG. 8D illustrates the effect of the window size on transition-time and decoding-accuracy when using the LF data for subjects 1 and 2,
- each trial included a single sentence, Before each block, the subjects were instructed to pay attention to one of the speakers (starting with the male) and to switch attention on each successive block. In order to test the intelligibility of the speech, after each trial (sentence) the subjects were presented with a transcription of the sentence of the attended speaker with one word missing. Subjects were instructed to type the missing word (the intelligibility task).
- 9B is a graph showing the MOS scores when the system was off and on, and illustrates that MOS measures were significantly improved when the system was on. Particularly, all but one subject reported a larger MOS when the system was on, with a median MOS of 3.87 (25 iJl percentile, 3.73; 75 ib percentile, 4.12) versus a median MOS of 3.5 (25th percentile, 2.9; 75th percentile, 3.8) when the system was off. The one subject who reported no reduction in listening effort had almost identical MOS scores for both system on (3.7) and system off (3.8).
- LF data was used, as opposed to the HF data, to be as comparable as possible with noninvasive technologies which typically only have access to low-frequency neural activity.
- a window size of 20 seconds was used in order to be consistent with the decoding strategies discussed herein. Whenever it was possible to correctly classify the attended speaker from the neural data for that subject, the output from the correct DNN was added to the mixture. However, if a mistake was made, and the subject's attentional focus was misclassified, the output from whichever DNN produced the largest normalized correlation was presented.
- the DNN output was added at a level of +12dB relative to the mixture, hi addition to obtaining measures of intelligibility and listening effort, a determination of the participants' overall preference was made. Subjects were informed before the experiment that they would have to report which half of the experiment requi red less effort to understand the attended speaker, and were reminded half way through. The experimentation also asked the users as to whether they would prefer to use the system if they knew that it was not perfect.
- an end-to-end system that incorporates a single channel automatic speech-separation process into the auditory attention-decoding (AAD) platform was described.
- the approaches discussed herein alleviate the spatial separation requirements of multi-channel approaches, but can also be used in tandem with beamforming methods for optimal source separation.
- the implementations described herein also amplify that speaker, resulting in an improvement in the subjective quality of the listening experience.
- the systems described herein can be used to realize hearing aid devices that can automatically and dynamically track a user's direction of attention, and amplify an attended speaker (and/or attenuate non-attended speaker(s)).
- neurological signals can be obtained through invasive ECoG recordings from neurological patients, which allows examining the upper bound of decoding speed and accuracy, determining the varying contributions of different auditory cortical areas, and separately examining attention decoding using low (l-8Hz) and high (70-150Hz) frequency neural data.
- neural data may be obtained also through non-invasive neural recordings.
- each DNN of the implementation may therefore be able to separate a trained target speaker from other unseen speakers.
- the proposed system can switch to default operation when no target speaker is detected. Generally, a new speaker can be easily added to the system, requiring only a small amount, of clean speech from the new target speaker (e.g., -20 minutes). Because the reconstruction process from neural data may be achieved using single speaker data training, adding a new speaker does not involve recording new neural responses for that speaker.
- a practical limitation for all processes intended for hearing aids (or other hearing augmentation devices) is that hardware constraints could limit the number of DNNs that could be housed inside a portable device.
- modem hearing aids are able to perform off-board computing by interfacing with a ceil phone, and specialized hardware is also becoming available for low-power neural network implementations.
- Another consideration is the fact that DNNs rely heavily on the data used to train them, Therefore, additional training may be required to separate speakers under different environmental conditions. Also, because people tend to involuntarily speak louder in noisy situations, which affects acoustic features such as pitch, rate and syllable duration, this would also need to be taken into account during training of the DNNs.
- This problem is particularly important when using neural signals with a lower signal to noise ratio (such as around the ear, or in ear EEG). It is possible that more elaborate decoding processes can be used to speed up decoding and provide a better trade-off between decoding-accuracy and transition-time.
- a deep learning framework also referred to as the attractor network
- the term attractor is motivated by the well-studied perceptual effects in human speech perception which suggest that the brain circuits create perceptual attractors (magnets) that warp the stimulus space so as to draw the sound that is closest to it, a phenomenon that is called Perceptual Magnet Effect.
- the proposed model works on the same principle by forming a reference point (attractor) for each source in the embedding space which draws all the T-F bins toward itself. Using the similarity between the embedded points and each attractor, a mask is estimated for each sources in the mixture. Since the mask is directly related to the attractor point, the proposed framework can potentially be extended to arbitrary number of sources without the permutation problem.
- the mask learning allows a very efficient end-to-end training scheme and highly reduces the computation complexity compared with the deep-clustering (DC) approach and the permutation invariant training (PIT) approach.
- DC deep-clustering
- PIT permutation invariant training
- T-F time-frequency
- PIT permutation invariant training
- DC is generally able to solve both permutation and output dimension problem to produce the state of the art separation performance.
- the main drawback of DC is its relative inefficiency to perform end-to-end mapping, because the objective function is the affinity between the sources in the embedded space and not the separated signals themselves.
- Minimizing the separation error is done with an unfolding clustering system and a second network, which is trained iteratively and stage-by-stage to ensure convergence.
- the PIT algorithm solves the permutation problem by pooling over all possible permutations for N mixing sources (jV! permutations), and using the permutation with lowest error to update the network.
- the PIT approach was shown to have comparable performance as DC.
- PIT approach suffers the output dimension mismatch problem because it assumes a fixed number of sources.
- PIT also suffers from its computation efficiency, where the prediction window has to be much shorter than context window due to the inconsistency of the permutation both across and within sample segments.
- a diagram of a proposed system architecture 1000 to implement a deep-attractor network approach is provided.
- the system 1000 receives as input a mixture signal 1002 (e.g., a multi-speaker combined signal) via an input sensor (e.g., a single or multiple microphones coupled to a device such as a hearing devices, or some other hearing augmentation device).
- the received input mixture signal 1002 is provided to a deep learning neural network 1010, which may be implemented, for example, as an LSTM layer.
- the neural network is trained to map the mixture sound X to a k dimensional embedding space (represented by the box 1020), such that it minimizes an objective function such as the following:
- the mask is estimated in the K dimensional embedding space of each T-F bin,
- Equation 7 The objective function in Equation 7 includes of three parts. During training, an embedding Vis first computed through a forward pass of the neural network for each given mixture. Then an attractor vector is estimated for each source using Equation 9, This can be done in several ways. The most straightforward method for attractor generation is to find the source centroid, as defined in Equation 9.
- a reconstruction mask is estimated (at box 1040) for each source by finding the similarity of each T-F bin in the embedding space to each of the attractor vectors A, where the similarity metric is defined in Equation 8.
- This particular metric uses the inner product followed by a sigmoid function which monotoiiicaily scales the masks between [0, 1], Intuitively, if an embedding of a T-F bin is closer to one attractor, then it means that it belongs to that source, and the resulting mask for that source will produce larger values for that T-F bin. Since the source separation masks for each TF bin should add up to one, particularly in difficult conditions, the sigmoid function (of Equation 8) can be replaced with softmax function:
- DANet deep attractor network
- DANet removes the stepwise pre-training process (required, for example, in the DC methodology) to allow end-to- end training.
- Another advantage of the DANet approach arises from flexibility in the number of sources in the mixture.
- Attractor points can be estimated using various processes other than the average-based approach.
- One possibility is to use weighted average. Since the attractors represent the sources' center of gravity, only the embeddings of the most salient T-F bins may be included, which leads to more robust estimation.
- This strategy can be investigated by using an amplitude threshold in the estimation of the attractor.
- a neural network model may also be used to pick the representative embedding for each source, an idea which shares similarities with encoder- decoder attention networks.
- two strategies can be applied to form the attractor points. The first is similar to the strategy used in DC, where the centers are found using post K-means process.
- the second strategy is based on the observation that the location of the attractors in the embedding space is relatively stable. This observation is shown in FIGS. 11 A-B, which includes, at FIG. 11 A, graphs showing an example of a mixture, the difference between two speakers in the mixture, and the separated spectrograms of the two speakers using a DANet implementation, and, at FIG. 1 IB, a graph showing locations of T-F bins in the embedding space, where each pair of dots corresponds to the attractor found for the two speakers in a given mixture.
- Equation 11 The objective function of DC is shown in Equation 11, where Y is the indicator function which is equivalent to a binary mask, and V is the learned embedding: [00100] Since 7 is orthogonal and constant for each mixture, by multiplying Y 1 and a normalizer to both terms, an objective function can be derived that is a special case
- Equation can be viewed as an averaging operation, where the
- the freedom of the network to form attractor points during the training allows the system to use the affinity between samples where no constraint is on the number of patterns found, therefore allowing the network to be independent of the nuraber of sources.
- the flexibility of the network in choosing the attractor points is helpful even in two-source separation problems, because the two sources may have very different structures.
- FIG. 12 providing a graph 1200 presenting the locations of attractor points in the embedding space, the proposed implementations trained in speaker separation tasks have ended up finding 2 attractor pairs (4 points in the embedding space). More particularly, each point in FIG. 12 corresponds to one of 10,000 mixture sounds, visualized using the first three principal, components. Two distinct attractor pairs (denoted as A l and A2) are visible.
- the procedure 1300 includes acquiring 1310 a mixture speech signal from multiple sources, the mixture signal defining a time-frequency embedding space.
- the mixture signal may be the combined sounds signal of the procedure 300, which includes multiple sounds signals from multiple sounds sources.
- the mixture speech signal may be obtained using a single audio sensor, such as a microphone, or multiple audio sensors.
- the procedure 1300 further includes determining 1320 respective reference points for each of the multiple sources, with the reference points representing locations of the sources in the embedding space for the mixture speech signal.
- the reference points may include attractor points in the embedding space.
- determining the respective reference points may include determining the respective reference points using a deep neural network,
- the procedure 1300 further includes deriving masks for the determined reference points, and extracting 1340 at least one of the multiple sources using at least one of the derived masks. Deriving the masks may include computing similarity between embedded time- frequency points within the embedding space and the determined respective reference points.
- the deep attractor network approach for speech separati on may be used in many different application.
- the DANet approach may be used to separate multi- speaker speech signals to implement attentional selection processing (i.e., enhance the speech signals of a speaker that the listener is determined to be listening to or focused on).
- the DANet approach may also be used in surveillance applications (to separate and record speech from a combined sound signal corresponding to different speakers), for virtual reality applications (reconstruct speaking objects from a pre-recorded source, and render the separated speech signals into a virtual audio, or audio-visual, scene), etc.
- a three speaker mixture dataset was constructed for a three speaker separation evaluation from the same WSJ set, which had 30 hour training data, 10 hours validation data, and 5 hours testing data, mixed with SNR at -5 ⁇ 5 dB. Each three speaker mixture was ensured to include both female and male speakers. All data was re-sampled to 8 kHz to reduce computational and memory costs. The log spectral magnitude was served as an input feature, computed using short-time Fourier transform (STFT) with 32 ms window length, 8 ms hop size, and the square root of harming window.
- STFT short-time Fourier transform
- the network contained four (4) bi-directional LSTM layers with 600 hidden units in each layer.
- the embedding dimension was set to 20, resulting in a fully-connected feedforward layer of 2580 hidden units (20 x 129) after the BLSTM layers.
- the input features were split into non-overlapping chunks of 100-frame length as the input to the network.
- An RMSprop procedure was used for training, with an exponential learning rate decaying strategy, where the learning rate started at 10 -4 and ended at 3 x 10 -6 ,
- the total number of epochs was set to be 150, and a cost function was used in Equation 7 on the validation set for early stopping. The criteria for early stopping is no decrease in the loss function on validation set for 10 epochs.
- a Deep Clustering (DC) network was constructed with the same configuration, which was used as the baseline.
- SDR signal-to-distortion ratio
- SAR signal-to-artifacts ratio
- SIR signal-to-interference ratio
- a neural network framework referred to as a deep attractor network for general source separation problem.
- the network forms attractor points in a high-dimensional embedding space of the signal, and the similarity between attractors and time-frequency embeddings are then converted into a soft separation mask.
- the proposed framework can perform end-to-end, real-time separation, and can work on different number of mixing sources. Evaluations and experimentations performed for the deep attractor network implementations for multi-speaker separation tasks confirmed its efficacy and potential for extension to general source separation problems.
- a controller system e.g., a processor-based controller system, a state machine, etc.
- a controller system e.g., a processor-based controller system, a state machine, etc.
- the various devices/systems e.g., hearing devices, or other types of hearing augmentation devices
- the computing system 1400 includes a computing-based device 1410 such as a personal computer, a specialized computing device, a controller circuit (implemented on a chip), and so forth, that typically includes a central controller 1412 (which may a programmable processor, such as a CPU).
- the system includes main memory, cache memory and bus interface circuits (not shown in FIG. 14).
- the computing / controller-based device 1410 may include a mass storage element 1414, such as a hard drive or flash drive associated with the computer system.
- the computing / controller system 1400 may further include a keyboard 1416, or keypad, or some other user input interface, and a monitor 1420, e.g., an LCD (liquid crystal display) monitor, that may be placed where a user can access them.
- the computing system 1400 may be incorporated within a system such as the various devices described herein.
- the computing-based device 1410 is configured to facilitate, for example, the implementation of operations for neural decoding of attention al selection in multi -speaker environments and/or for deep attractor networks.
- the storage device 1414 may thus include a computer program product that when executed on the computing / controller-based device 1410 causes the computing-based device to perform operations to facilitate the implementation of procedures and operations described herein.
- the computing / controller-based device may further include peripheral devices to enable input/output functionality.
- peripheral devices may include, for example, a CD-ROM drive and/or flash drive (e.g., a removable flash drive), or a network connection (e.g., implemented using a USB port and/or a wireless transceiver), for downloading related content to the connected system.
- peripheral devices may also be used for downloading software containing computer instaictions to allow general operation of the respective system/device.
- special purpose logic circuitry e.g., an FPGA (field programmable gate array), an ASIC
- the computing- based device 1410 may include an operating system, e.g., Windows XP® Microsoft Corporation operating system, Ubuntu operating system, etc,
- Computer programs include machine instructions for a programmable processor, and may be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language.
- 'machine-readable medium ' refers to any non-transitory computer program product, apparatus and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a non-transitory machine-readable medium that receives machine instructions as a machine-readable signal.
- PLDs Programmable Logic Devices
- any suitable computer readable media can be used for storing instructions for performing the processes / operations / procedures described herein.
- computer readable media can be transitory or non-transitory.
- non-transitory computer readable media can include media such as magnetic media (such as hard disks, floppy disks, etc.), optical media (such as compact discs, digital video discs, Blu-ray discs, etc.), semiconductor media (such as flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read only Memory (EEPROM), etc.), any suitable media that is not fleeting or not devoid of any semblance of permanence during transmission, and/or any suitable tangible media.
- transitory computer readable media can include signals on networks, in wires, conductors, optical fibers, circuits, any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and/or any suitable intangible media.
- substantially as used herein when referring to a measurable value such as an amount, a temporal duration, a physical attribute (such as frequency), and the like, also encompasses variations of ⁇ 20% or ⁇ 10%, ⁇ 5%, or +0.1% from the specified value, as such variations are appropriate in the context of the systems, devices, circuits, methods, and other implementations described herein.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Human Computer Interaction (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Neurosurgery (AREA)
- Multimedia (AREA)
- Acoustics & Sound (AREA)
- Signal Processing (AREA)
- Neurology (AREA)
- Dermatology (AREA)
- Automation & Control Theory (AREA)
- Fuzzy Systems (AREA)
- Otolaryngology (AREA)
- Quality & Reliability (AREA)
- Circuit For Audible Band Transducer (AREA)
Abstract
Disclosed are devices, systems, apparatus, methods, products, and other implementations, including a method that includes obtaining, by a device, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources. The method further includes obtaining, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, selecting one of the plurality of separated signals based on the obtained neural signals for the person, and processing the selected one of the plurality of separated signals.
Description
NEURAL DECODING OF ATTENTIONAL SELECTION IN MULTI-SPEAKER
ENVIRONMENTS
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims the benefit of, and priority to, U.S. Provisional Patent Application Serial Nos. 62/480,005, entitled "DEEP ATTRACTOR NETWORK FOR SINGLE CHANNEL SPEECH SEPARATION" and filed March 31, 2017, 62/429,549, entitled "NEURAL
DECODING OF ATTENTIONAL SELECTION IN MULTI-SPEAKER ENVIRONMENTS WITHOUT ACCESS TO SEPARATED SOURCES" and filed December 2, 2016, and
62/349,976, entitled "NEURAL DECODING OF ATTENTIONAL SELECTION IN MULTI- SPEAKER ENVIRONMENTS WITHOUT ACCESS TO SEPARATED SOURCES" and filed June 14, 2016, the contents of all which are incorporated herein by reference in their entireties.
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH
[0002] This invention was made with government support under Grant No. DC014279 awarded by the National Institutes of Health (NIH), and under NSF Career Award No. 1 SSS079 granted by the National Science Foundation. The government has certain rights in the invention.
BACKGROUND
[0003] In a natural auditory environment, most people can generally attend to (i.e., focus on) a particular speaker out of many. However, this task is challenging for persons suffering from peripheral and central auditory pathway disorders.
SUMMARY
[0004] In some variations, a method is provided that includes obtaining, by a device, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals
that each contains signals corresponding to different groups of the multiple sound sources. The method further includes obtaining, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, selecting one of the plurality of separated signals based on the obtained neural signals for the person, and processing the selected one of the plurality of separated signals.
[0005] Embodiments of the method may include at least some of the features described in the present disclosure, including one or more of the following features,
[0006] The method may further include generating a sound spectrogram from the combined sound signal, and applying the neural-network-based speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
[0007] Selecting the one of the plurality of separated sound signals based on the obtained neural signals for the person may include generating an attended speaker spectrogram based on the neural signals for the person, comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms, and transforming the selected one of the multiple resultant speaker spectrograms into an acoustic signal.
[0008] Comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms may include comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
[0009] Obtaining the neural signals for the person may include obtaining one or more of, for example, electrocorticography (ECoG) signals for the person, neural measurements via a noninvasive scalp, and/or in-ear EEG recordings.
[0010] Applying the neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the combined sound signal.
[0011] Processing the selected one of the plurality of separated sound signals may include one or more of, for example, amplifying the selected one of the plurality of separated signal s, and/or attenuating at least one non-selected signal from the plurality of separated signals.
[0012] Obtaining the combined sound signal for the multiple sound sources may include receiving the combined sound signal for the multiple sound sources at a single microphone coupled to the device.
[0013] The combined sound signal may define a time-frequency embedding space. Applying the neural-network-based speech-separation processing to the combined sound signal may include determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sound sources in the embedding space for the combined sound signal, deriving masks for the determined reference points, and extracting at least one of the multiple sound sources using at least one of the derived masks.
[0014] Determining the respective reference points may include determining the respective reference points using a deep neural network.
[0015] Deriving the masks may include computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points.
[0016] The reference points may include attractor points in the embedding space.
[0017] In some variations, a system is provided that includes at least one microphone to obtain a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, one or more neural sensors to obtain neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, and a controller coupled to the at least one microphone and the one or more neural sensors. The controller is configured to apply neural -network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
[0018] In some variations, non-transitory computer readable media are provided, thai are programmed with instructions, executable on a processor, to obtain, by a device comprising the processor, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and apply, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of
separated signals that each contains signals corresponding to different groups of the multiple sound sources. The instructions further include one or more instruction, executable on the processor, to further obtain, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
[0019] In some variations, an apparatus is provided that includes means for obtaining a combined sound signal for signals combined from multiple sound sources in an area in which a person is located, and means for applying neural -network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources. The apparatus further includes means for obtaining neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, means for selecting one of the plurality of separated signals based on the obtained neural signals for the person, and means for processing the selected one of the plurality of separated signals.
[0020] Embodiments of the system, the computer-readable media, and the apparatus may include at least some of the features described in the present disclosure, including at least some of the features described above in relation to the method.
[0021] Details of one or more implementations are set forth in the accompanying drawings and in the description below. Further features, aspects, and advantages will become apparent from the description, the drawings, and the claims.
BRIEF DESCRIPTION OF THE DRAWINGS
[0022] These and other aspects will now be described in detail with reference to the following drawings.
[0023] FIG. 1 is a diagram of a system to process (e.g., separate) combined sound signals (e.g., for attend onal selection applications).
[0024] FIG. 2 is another example implementation of a system to process (e.g., separate) sound signals,
[0025] FIG. 3 is a flowchart of an example procedure for attentional selection of a speaker in a multi-speaker environm ent.
[0026] FIG. 4 is a graph providing an analysis of deep neural network (DNN) output correlation.
[0027] FIG. 5 includes graphs illustrating reconstruction accuracy for neural correlation analysis.
[0028] FIG. 6A includes graphs showing the proportion of segments in which the attentional focus of subjects could be correctly determined using HF data and LF data.
[0029] FIG. 6B includes graphs presenting attention decoding index (ADI) measures per electrode.
[0030] FIG. 7 is a graph showing results of an analysis for dynamic switching of attention for an example subject.
[0031] FIGS. 8A-D are graphs showing further performance results in relation to dynamic switching of attention
[0032] FIGS, 9A-B are graphs illustrating speaker separation quality results.
[0033] FIG. 10 is a diagram of a proposed system architecture to implement a deep-attractor network (DANet) approach.
[0034] FIG. 11 A includes graphs showing an example of a mixture, the difference between two speakers in the mixture, and the separated spectrograms of the two speakers using a DANet implementation.
[0035] FIG. 1 IB is a graph showing locations of T-F bins in an embedding space. [0036] FIG. 12 is a graph presenting the locations of attractor points in the embedding space, [0037] FIG. 13 is a flowchart of an example deep attractor signal separation procedure. [0038] FIG. 14 is a schematic diagram of a generic controller system. [0039] Like reference symbols in the various drawings indicate like elements.
DETAILED DESCRIPTION
[0040] Described herein are systems, devices, apparatus, methods, computer program products, media, and other implementations, to process sound signals / data in environments that contain multiple sounds sources, based on neural signals / data from a person receiving the multi-source sound signals. In sorae erabodiraents, procedures to decode which speaker a person is attending to are implemented by monitoring their neural activity via both invasive and non-invasive electrophysiological recordings. In such embodiments, attention-aware brain-computer interfaces (BCIs) may be used to control smart bearable devices capable of selectively processing (e.g., amplifying) one speaker and/or suppressing other speakers in crowded environments.
[0041] In some attention decoding procedures, explicit access to the isolated sound sources in the environment may be required, which may be difficult to achieve. Techniques such as beamforming may be used in an attempt to isolate different speakers. However, these techniques may not be entirely adequate when there is little (or no) spatial separation between speakers. Also, the techniques' reliance on multiple microphones has drawbacks with regards to hardware complexity and user comfort. Thus, in implementations described herein, these challenges are addressed by using a single microphone, or by combining the beamforming techniques / methods with the proposed implementations, to implement automatic speech separation processing in which, for example, neural-network-based processing, such as deep neural network (DNN) models, are used to separate the sound sources. Generally, DNN refers to neural networks that include multiple hidden layers between the input and output layers of the networks. In DNN's that are used herein for speech separation processing, the output of the speech separation processes can be used to decode the attentional state of the listener, and subsequently, to amplify the attended source. For example, in some embodiments, a speech separation procedure separates out the speakers in a crowd, and compares to a signal reconstructed from the brain. The speech separation procedure can next amplify (or otherwise filter) that speaker source compared to other sources.
[0042] Accordingly, implementations described herein include a method that comprises obtaining, by a device (e.g., a hearing device, a hearing enhancement device to allow better hearing in a very noisy environment, a noise cancellation device, a virtual reality device, and/or
other types of hearing augmentation devices), a combined sound signal for sound combined from multiple sound sources in an area in which a person is located. The method further includes applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, and obtaining, by the device, neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to. The method additionally includes selecting one of the plurality of separated signals based on the obtained neural signals for the person, and processing the selected one of the plurality of separated signals. In some embodiments, obtaining the neural signals for the person may include obtaining one or more of, for example, electrocorticography (ECoG) signals for the person, neural measurements via a non-invasive scalp, and/or in-ear EEG recordings. Applying the speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the multiple sound sources based on the combined sound signal. Processing the selected one of the plurality of separated sound signals may include one or more of, for example, amplifying the selected one of the plurality of separated signals, and/or attenuating at least one non-selected signal from the separated sound signals.
[0043] In some embodiments, separation processes to separate the multi-source audio / sound signals may be realized as a deep learning framework for single channel speech separation by creating attractor points in high dimensional embedding space of the acoustic signals, which pull together the time-frequency bins corresponding to each source. Attractor points are created, in some embodiments, by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings.
[0044] Thus, in embodiments of the deep attractor network approach, the combined signal may be a mixture speech signal defining a time-frequency embedding space. Applying the speech- separation processing to the combined sound signal may include, in such embodiments, determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sources in the embedding space for the mixture
speech signal, deriving masks for the determined reference points, and extracting at least one of the multiple sources using at least one of the derived masks. The reference points may include attractor points in the embedding space. Determining the respective reference points may include determining the respective reference points using a deep neural network. Deriving the masks may include computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points.
Attentional Selection in Multi-Speaker Environments
[0045] With reference now to FIG. I, a diagram for an example system 100 to process sound signals is shown. The system 100 (which may be implemented as a hearing device to facilitate hearing for a person 104, or may be implemented as a noise cancellation device, or any other type of a hearing augmentation device) includes a microphone 110 to receive a combined sound signal 102 combining sound signals from multiple sound sources. In some embodiments, a controller (e.g., a computing-based device, which may be incorporated into the hearing device, or some other hearing augmentation device, used by the person 104) implements one or more sound-processing procedures 120 (which may each correspond to a separate DNN, although, in some embodiments, a single neural network unit may be configured to implement multiple separate DNN's) to process the combined (aggregated) sound signal 102 to separate the combined sound signal into a plurality of separate signals 122a-n (which may include sound signals, or may include signal representative and/or derived from sound signals). For example, in some embodiments, the one or more procedures 120 implemented by the system's controller may include a speech-separation process based on a deep-neural network.
[0046] The system 100 also includes one or more sensors to obtain / measure neural signals / data 106 from the person 104 that are indicative of the sound source (e.g., one of multipl e speakers in the vicinity of the person 104) the person 104 is attentive to. Based on the neural signals / data obtained for the person 104, one or more of the plurality of separated signals derived by the one or more procedures 120 is selected. In the example of FIG. 1, the neural signals obtained for the person 104 may indicate that the user is attentive to the sound source SP2, and, therefore, separated sound signal 122b (corresponding to the source SP2 in FIG. 1) is selected for further processing. The selected separated sound signal can then be processed to, for
example, amplify, or otherwise perform signal enhancing filtering on the signal. Additionally and/or alternatively, the non-selected signals (i.e., the signals other than the sound signal 122b) may be processed to attenuate or suppress them (or to perform some other filtering operation thereon),
[0047] FIG. 2 depicts another example implementation of a system 200 to process (e.g., separate) sound signals (e.g., in a closed-loop). As shown, two speakers, namely, Spkl and Spk2, are mixed together, as illustrated in a first stage 210 of the system 200. In some embodiments, the first stage 210 includes various signal recording and pre-processing operations performed on acquired audio signals. In example implementations that were realized and evaluated, as will be more particularly described below, some source speakers were recorded using Tucker Davis Technologies (TDT©) hardware and sampled at 2441Hz, while other source speakers were recorded using Xltek® hardware and sampled at 500Hz. Further processing operations were performed offline. In the implementations that were realized and tested herein, filters were realized using Matlah's® Filter Design Toolbox, and were used in both a forwards and backwards direction to remove phase distortion. In the example using TDT acquired data, the data was resampled to
and a 1 st order Butterworth highpass filter with a cut-off frequency at IHz was used to remove DC drift. The data was subsequently re-referenced using a common average scheme, with the effect of noisy electrodes mitigated by using a trimmed mean approach (e.g., removing the top and bottom 10% of electrodes at each sample point). These example implementations also included removing line noise at 60Hz and its harmonics (up to 240Hz) using 2nd order IIR notch filters with a bandwidth of IHz. A period of silence was recorded before each experiment, and the corresponding data was normalized by subtracting the mean and dividing by the standard deviation of the pre-stimulus period. These implementations also included filtering data into two frequency bands known to be responsive to speech, namely, the high-frequency (HF) gamma band (70-150Hz), and the low-frequency (LF) delta and theta bands (l-8Hz). The power of the high-frequency band and the phase of the low-frequency band are modulated by speech, In order to obtain the power of the HF band, the data was first filtered into 8 frequency bands between 70 and 150Hz, using, for example, Chebyshev Type 2 filters, each with a bandwidth of 10Hz. The power analytic amplitude of each band could then be derived using a Hilbert transform. The average of all 8 frequency bands can then be used to represent the total HF power. To obtain the LF data, an 8th order (4 sections) Butterworth low-
pass filter was applied to the data with a corner frequency of, for example, 8Hz, and 20dB attention at 12Hz,
[0048] With continued reference to FIG. 2, signal separation processing is applied to the combined (mixed) audio signal (which may have been processed based on, for example, processing operation such as those described in relation to stage 210) at stage 220. In some embodiments, speech separation processing is based on a deep neural network (DNN) approach. After obtaining the mixed / combined audio signal, a spectrogram of the mixture is obtained. This spectrogram is then fed to each of several DNNs, each trained to separate a specific speaker from a mixture (DNN Spkl to DNN SpkN). At the same time, a user may be attending to one of the speakers (in this case, Spkl). A spectrogram of this speaker is reconstructed from the neural recordings of the user. This reconstruction is then compared with the outputs of each of the DNNs using a normalized correlation analysis (as shown at stage 230) in order to select the appropriate spectrogram, which is then converted into an acoustic waveform and added to the mixture so as to amplify the attended speaker (and/or attenuate the signals corresponding to the non-attended speakers).
[0049] More particularly, In order to automatically separate each speaker from the mixture, a single-channel speech separation that utilizes, in some embodiments, a class of DNNs known as long short-term memory (LSTM) DNNs is employed. Each DNN is trained to separate one specific speaker from arbitrary mixtures. In an experiment, two speakers (Spkl F and Spk2M) were presented to each subject. However, a system configured to work in a real -world situation would generally contain multiple DNNs, each trained to separate specific speakers, any of whom may or may not be present in the environment. In the example system of FIG. 2, 4 DNNs were trained to separate 4 speakers (two female speakers and two male speakers), hereafter referred to as SpklF, Spk2M, Spk3F, and Spk4M. In an example, ail speakers were native American English speakers. In order to be used by the DNNs, the speech waveforms were converted, in the example experiment, into 100-dimensional Mel-frequency spectrograms. The goal was then to obtain an estimate S' of a clean target spectrogram S from a mixture M. To do so, a soft mask Y' is learned and applied to the mixture to mask the interfering speech. The squared Euclidian distance between the masked spectrogram and the clean target spectrogram is treated as the error in order to generate the gradient that is back propagated through the DNN to update the parameters. An objective function that may be used is:
[0050] In the experimentation conducted for the implementation s if the system 200, the input to the DNNs was the logarithm (base 10) of the spectrograms, normalized so that each frequency band had zero mean and unit variance. Each DNN, in the implementations realized and tested, contained four (4) layers with 300 nodes each, followed by a single layer containing 100 nodes with logistic activation to output a spectrogram. An acoustic waveform was generated by combining the spectrogram that was output by the DNN with the phase of the original mixture (e.g., inverting, in some embodiments, the spectrogram with the phase of the original mixture). In the implementations that were realized and tested, twenty (20) minutes of speech from the target speakers, and approximately five (5) hours of speech from interfering speakers were used. A target speaker sample was generally mixed with one interfering speaker, and both were mixed into the same channel and with the same root mean squared (RMS) intensity. Unseen utterances were used for testing (for both the target and interfering speakers). To ensure generalization, each of the DNNs did not receive training sample data for the other target speakers during training. That is, in the implementations realized and evaluated herein, a DNN trained to separate one speaker (e.g., SpklF) was not exposed to training data for other speakers,
[0051] As will be discussed in greater detail below, in some embodiments, other types of speech separation processing may be applied to the mixed / combined signal. As noted, one speech separation approach that may be implemented is a deep attractor network (DANet) approach for separating multiple sound sources (e.g. speakers) from a single channel recording. Under this approach, the a deep learning framework for single channel speech separation is used that is configured to create attractor points in high dimensional embedding space of the acoustic signals, which pull together the time-frequency bins corresponding to each source. Attractor points maybe created by finding the centroids of the sources in the embedding space, which are
subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. Advantages of the DANet approach is that it implements an end-to-end training, and it does not depend on the number of sources in the mixture. Two strategies that can be used with respect to the DANet implementations are the K-means strategy and the fixed-attractor-
points strategy, where the latter requires no post-processing and can be implemented in real-time. In example system implementations of the DANet approach, the system was configured to output several waveforms, each corresponding to one sound source in the mixture, in response to an input mixed waveform, The processing operations of such a system implementation may include acquiring mixture speech signal with a (single) sensor, processing the signal to estimate a set of features, processing the features by a deep neural network (DNN) to assign a descriptor for each element in the signal, estimating the attribution of each element according to the physical properties of the descriptors, and recovering the target sources according to the estimation.
Example embodiments based on the DANet approach can be incorporated into online
applications or mobile devices to allow fast and accurate signal separation, and can also be deployed into hearing aids or earphones to be a front-end solution of source separation.
[0052] While the embodiments described herein may use multiple possible speech separation approaches, for ease of illustration and expl anation, the approach using multiple DNN's trained for specific respective speakers (i.e., the approach illustrated in FIGS, 1 and 2) will be referenced. However, as noted, the identification of speakers a listener is attending to can be performed using other speech separation approaches, including the DANet approach.
[0053] Thus, with continued reference to FIG. 2 (and similarly to FIG. 1), selection of the speaker, whose speech data is to be provided to the target user, is performed through use of a speaker selection stage 230 and a stimulus reconstruction stage 240 of the system 200. The signals corresponding to the selected speaker are then processed (e.g., amplified) using the filtering processes implemented at stage 250, Additionally and/or alternatively, signals corresponding to non-selected speakers mat be processed to, for example, attenuate those signals. In some embodiments, the stimulus-reconstruction procedure implemented by the stage 240 applies a spatiotemporal filter (decoder) to neural recordings to reconstruct an estimate of the spectrogram of the attended speaker, A 100-dimensional log Mel-frequency spectrograms maybe downsampled by a factor of 10, resulting in ten (10) frequency bands used for neural recordings. Each decoder can be trained using the data from the single- speaker (S-S)
experiment.
[0054] Electrodes may be chosen if they are significantly more responsive to speech than to silence. For the high gamma (HG) data, this means that the average power during speech is
greater than during silence. For the low frequency (LF) data, electrodes may be chosen if the average power is either significantly greater or lower during speech than during silence. To perform statistical analyses, the data can be segmented into SOOms chunks and divided into two categories; speech and silence. Significance may then be determined using an unpaired t-test (false discovery rate (FDR) corrected, q < 0.05). This results in varying numbers of electrodes retained for each frequency band and each subject, The decoders may be trained using time-lags from -400 to 0 ms.
[0055] Determining to whom the subject user is attending (i.e., which speaker the target user is trying to listen to) may be implemented through a correlation analysis, e.g., using Pearson's r- value (e.g., the linear correlation measure between two variables). Typically, whichever spectrogram has the largest correlation with the reconstructed spectrogram is considered to be the attended speaker. However, in embodiments in which four (4) DNNs are used, each trained to separate a speaker that may or may not be present in the combined signal (mixture), the analysis becomes slightly more complex. Because the DNNs that do not see their designated speakers in the mixture are likely to output a signal very close to the mixture, it may be necessary to normalize the correlation values with respect to the mixture. This is because the correlation between the reconstructed spectrograms and the mixture can be very large.
[0056] In the continued discussion of correlation analysis processing, the following terminology and notations will be used. The
decoder configured to identify the
specific speaker) is referred to as 5DNNK, the spectrogram of the mixture signal is represented as 5MIX, and the reconstructed spectrogram (from the neural responses) as SRECON. in order to emphasize large correlations, a Fisher transformation (inverse hyperbolic tangent) is applied to each
[0057] The normalization procedure involves five stages. First, the correlation between SRECON and each which is denoted as pik is derived according to:
where
is Pearson's correlation between the variables x and y, and arctanh is the inverse hyperbolic tangent function.
[0058] Next, the correlation between
and the difference between and SMX, which is
relatively small if it is outputting the attended speaker (because subtracting the attended spectrogram from the mixture will only leave behind portions of the unattended spectrogram), and should be relatively large if it outputs the unattended speaker (similarly, because only potions of the attended speaker will be left). Therefore, taking the difference of and
where and are the mean and standard deviation of the distribution of a scores from all DNNs.
[0061] The next neural correlation processing stage includes subtracting the correlation between and adding the constant 1, providing a resultant normalized correlation value,
[0062] The above normalization operation effectively penalizes a DNN that is simply outputting the mixture rather than separating the speakers. This could occur if a DNN's trained speaker was not in the mixture. The addition of the constant 1 is used to make the final result more intuitive, as otherwise the values would typically be less than zero.
[0063] In experimentations and testing of the neural correlation processing realized as part of the implementations described herein, the lowest two frequency bands (-50 - 200Hz) of the 10
frequency bands in the downsampled spectrograms were excluded to avoid bias towards male speakers whose fundamental frequency occupied the exclusion range region. Correlation values derived may thus correspond to the average of the r-values obtained across the remaining eight (8) non-excluded frequency bands.
[0064] In order to obtain a measure of the ability of the implemented systems described herein to determine the attended speaker from neural recordings, the reconstructed spectrogram is first segmented (from the multi-speaker experiments) into, for example, 20-second bins. In an example implementation, in which four (4) DNNs are trained to separate two female speakers (denoted as Spkl F and Spk3F) and two male speakers (Spk2M and Spk4M) from random mixtures, four normalized correlation values are produced by the DNN's for each segment (the normalized correlation values are denoted
When, for example, SpklF and Spk2M are the only speakers that are presented to the subject (the listener), it is expected that PI. /-and P2;;/ would be the largest, depending on whom the subject was attending to. If there are only two possible correlation values to choose from, a segment could be considered correctly decoded if the attended speaker produced the largest correlation with the reconstructed spectrogram. However, when there are multiple values to choose from, it is important to take into account any bias for a particular speaker. This is of particular importance when using intracranial data, because it is possible that some electrodes could be tuned to speaker-specific features, and respond to those features regardless of the attentional focus of the subject. To take into account any such potential bias, the Attention Decoding Index (ADI) metric is defined as the proportion of the number of correct hits minus the number of false positives, for target speakers. The ADI metric may be computed as:
respectively, and
is the number of segments. The resultant metric is bounded between [-1,1]. In some of the experimentations and evaluation performed for the systems implemented herein, chance and significant ADI were determined by randomly shuffling the reconstructed spectrograms with respect to the DNN outputs 100 times for each subject. This allowed finding the actual chance performance and comparing that to the performance of the system (to test the signifi cance of the result), A null distribution of the ADI was then obtained using the same
normalized correlation analysis described above. The resulting mean ÷/- SD performance was 0 +/- 0.15 for both HF and LF data. Significant performance was therefore determined to be 0.45 (three times the standard deviation). For comparison, the ADI that would be achieved using the clean (ideal) spectrograms of Spkl F and Spk2M was also computed. In that ideal situation, it was assumed that the DNNs trained on Spk3F and Spk4M outputted the mixture.
[0065] In some of the experimentations and evaluations conducted, in order to simulate a dynamic scenario in which a subject (a listener) was switching attention, the neural data was artificially divided and concatenated into consecutive segments in which subjects were attending to either speaker. Specifically, the data was divided into ten (10) segments, each lasting 60 seconds. The subjects attended to the male speaker for the first segment. To assess the ability to track the attentional focus of each subject, a sliding window approach was used in which correlation values were obtained every second over a specified window. Window sizes ranging from 5 to 30 seconds (in 5 second increments for 6 window sizes in total) were used. Larger windows should lead to more consistent (i.e., less noisy) correlation values and provide a better estimate of the attended speaker. However, they should also be slower at detecting a switch in attention.
[0066] Thus, implementations described herein include a system comprising at least one microphone to obtain a combined sound signal for sound combined from multiple sound sources in an area in which a person is located, one or more neural sensors to obtain neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to, and a controller coupled to the at least one microphone and the one or more neural sensors. The controller is configured to apply speech-separation processing (e.g., neural-net processing such as deep neural network, or DNN) to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources, select one of the plurality of separated signals based on the obtained neural signals for the person, and process the selected one of the plurality of separated signals.
[0067] With reference now to FIG. 3, a flowchart of an example procedure 300 for attentional selection of a speaker in a multi-speaker environment. The procedure 300 includes obtaining 310, by a device, a combined or mixed sound signal for signals combined from multiple sound
sources in an area in which a person is located. As noted, the device of procedure 300 (that performs that various operation of the procedure 300) may include one or more of, for example, a hearing device, a hearing enhancement device (to allow better hearing in a very noisy environment), a noise cancellation device, a virtual reality device, and/or other types of hearing augmentation devices. In some embodiments, the combined signal may be obtained through use of a single receiving sensor (e.g., a single microphone), although the procedure 300 may be performed using multiple receiving sensor which may be part of the device implementing the procedure 300.
[0068] As further shown in FIG. 3, the procedure 300 includes applying 320, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources. Applying the neural-network- based speech-separation processing to the combined sound signal from the multiple sound sources may include providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the multiple sound sources based on the combined sound signal. In some embodiments, the DNN may include a plurality of individual DNN's (e.g., long short-term memory (LSTM) DNN's) that are each configured to separate one respective specific speaker from an arbitrary mixture of signals. Configuring the individual DNN's to separate their respective specific speakers may be realized through an initial training of the DNN's using speech data from the specific speaker, and speech data from interfering speakers (e.g., a combined / mixed speech signals). In some embodiments, speech separation processing may be applied to resultant data derived from the raw combined speech data (signals). For example, the speech separation processing may be applied to sound spectrogram data to derive multiple resultant speaker spectrograms. In such embodiments, the procedure 300 may further include generating a sound spectrogram from the combined sound signal, and applying the speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
[0069] As noted, in some embodiments, other speech separation approaches, such as the deep attractor network (DANet) approach, more particularly discussed below, may be used for the speech separating processing performed on the combined speech data, or performed on some resultant data derived from the combined speech data. Thus, in implementations to realize the
deep attractor network approach, the combined sound signal may define a time-frequency embedding space, and applying the speech-separation processing to the combined sound signal may include determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sources in the embedding space for the combined sound signal, deriving masks for the determined reference points, and extracting at least one of the multiple sound sources using at least one of the derived masks. In such embodiments, determining the respective reference points may include determining the respective reference points using a deep neural network. Deriving such masks may include computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points. The reference points may include attractor points in the embedding space.
[0070] While the speech-separation processes described herein are discussed primarily in relation to attentional selection application (e.g., to enhance the speech of a speaker, from a plurality of speaker, that the listener is focusing on), the speech-separation processes described herein may be utilized for other application. For example, the speech separation processes may¬ be used in surveillance application (to separate and record speech from a combined sound signal corresponding to different speakers). In other words, the speech separation processes described herein may be used to automatically analyze and transcribe survilance audio, which can be hundreds of hours long. The speech separation processes can be used as a preprocessing step, where the target speech sounds are first separated from other sources that might have been present during the recording, and the cleaned speech signal can be passed to speech recognition implementation. In another example, the speech separation processes described herein may be used for virtual reality applications, in which speech separation techniques discussed are used to reconstruct speaking objects from a pre-recorded source (and render the separated speech signals into a virtual audio (or audio-visual) scene). One of the challenges in 3D virtual audio is that rendering an acoustic scene requires separate audio tracks for individual sources, so that the sound sources can be processed separately and placed in different positions relative to the listener. When such separate audio tracks are not available, the speech processing procedures described herein can be used to first separate all the sources in the recording, and subsequently re-syntehsize them with desired location and attribute.
[0071] With continued reference to FIG. 3, the procedure 300 also includes obtaining 330, by the device, neural signals for the person, with the neural signals being indicative of one or more of the multiple sound sources the person is attentive to. Obtaining the neural signals for the person may include obtaining one or more of, for example, eiectrocorticography (ECoG) signals for the person, neural measurements via a non-invasive scalp, and/or in-ear EEG recordings.
[0072] Using the neural signals obtained for the person (the listener), the procedure 300 includes selecting 340 one of the plurality of separated signals based on the obtained neural signals for the person. As noted, in some embodiments, selecting the one of the plurality of separated signals may include generating an attended speaker spectrogram based on the neural signals for the person, comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms, and transforming the selected one of the multiple resultant speaker spectrograms into an acoustic signal. In such embodiments, comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms may include comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
[0073] The procedure 300 further includes processing 350 the selected one of the plurality of separated signals. For example, in some embodiments, processing the selected one of the plurality of separated sound signals may include one or more of amplifying the selected one of the plurality of separated signals, and/or attenuating at least one non-selected signal from the separated signals.
[0074] The implementations described herein were tested to evaluate their performance and efficacy. FIG. 4 includes a graph 400 providing an analy sis of the DNN output correlation. To test the performance of the DNNs at single-channel speaker-separation, multiple mixtures of random speakers were created, with four (4) target speakers (2 male and 2 female), and with each mixture being passed through four DNNs that were each pre-trained to separate one of the target speakers. Performance was measured by obtaining the correlation (r-value) between the output of each DNN and the spectrogram of the clean target speaker. The results from the four networks are split into two by averaging the r-values obtained from the networks that were trained to separate a femal e speakers (left porti on 410 of the graph 400 ) or male speakers (right portion 420 of the graph 400). The bars 412 and 422 show the results when a DNN was
presented with a mixture containing its designated pre-trained speaker (trained target), while the bars 414, 416, 424, and 426 are the results when the mixtures contained an undesignated speaker (i.e., an untrained target, with the bars 414 and 424 corresponding to female speakers, and 416 and 426 corresponding to make speakers), The dotted line 430 shows the average correlation between the raw mixture and the clean target speaker. The results illustrated in FIG. 4 show that, as expected, the networks could not separate undesignated speakers from the mixture, but their outputs were slightly more similar to the target speaker when that speaker was the same gender as the network's designated speaker. This result is likely because of the characteristic differences between male and female speakers (e.g., pitch, spectral envelope).
[0075] To determine which speaker a subject was attending to, a neural correlation analysis was performed to compare the reconstructed spectrograms (from the neural data) with the output of each DNN. FIG. 5 includes graphs 500, 510, 520, and 530, illustrating reconstruction accuracy for the neural correlation analysis. The left part of FIG. 5, including the graphs 500 and 510, provides the reconstruction accuracy in terras of raw r-values, and shows the average correlations between the reconstructed spectrograms and the outputs of the DNNs for each subject and neural frequency band (where each subject is represented by a dot of a different shade). Because, in the experiments conducted, the subjects alternated their attention between two speakers, the r-values labeled as attended and unattended come from the DNNs trained on SpklF and Spk:2M, The r- values labeled as undesignated come from the DNNs that were trained for Spk3F and Spk4M (as used herein, "undesignated" means that the DNN's in question were not trained to separate either speakers in the mixture that the listener actually listened to). Although the attended r-values are typically higher than the unattended r-values, there is also a high correlation between the reconstructed spectrograms and the mixture. Because the DNNs that do not see their designated speaker usually output spectrograms similar to the mixture, the undesignated correlation values were also relatively high.
[0076] In some embodiments, to improve the reconstruction analysis, the r-values are normalized with respect to the mixture. An example normalization solution involves three key steps; (i) incorporating the mixture into the correlation analysis (in accordance with, for example, Equations 2 and 3 above), (ii) performing a test-normalization (t-norm) to equalize the outputs of each DNN with respect to each other (in accordance with, for example, Equation 4, above), and (iii) subtracting the correlation between the DNN output and the mixture (in accordance with, for
example, Equation 5). The graphs 520 and 530 at the right portion of FIG. 5 provide the correlation analysis results in terms of normalized r-values, and indicate thai after applying the normalization processes, the attended correlation values were far higher than either the unattended or undesignated r-values, which allows adequate decoding of the attended speaker.
[0077] Given that the normalized correlation values differentiated between attended, unattended, and undesignated r-values, it is possible to decode the attentional focus of the subjects using the DNN outputs. After segmenting, in some implementations, the data into 20-second chunks, a segment was labeled as a correct hit (CH) if the normalized correlation between the
reconstructed spectrogram and the output of the DNN that was trained to separate the attended speaker was the highest of all 4 DNN outputs. A segment was considered to be a false positive if the correlation with the unattended speaker was higher than all others. The Attention Decoding Index (ADI) is defined as the proportion of segments correctly decoded minus the proportion of false positives. For comparison, the ADI obtained is also calculated using the ideal (clean) spectrograms of SpklF and Spk2M. FIG. 6A includes graphs 610 and 620 showing the proportion of segments (20s) in which the attentional focus of each subject could be correctly- determined using HF data (the graph 610 at the left part of FIG. 6 A) and LF data (the graph 620 at right part of FIG. 6A). The attentional focus of a subject was considered to be successfully decoded if the ADI was greater than the 0.45 threshold (indicated by the lines 612 and 622) calculated using a random shuffle of the data. Of the 6 subjects who participated in an experimental study, the attentional focus of subjects 1, 2, and 3 could be decoded using HF (70- 150Hz) data, and subjects 1, 2 and 5 using LF (l-8Hz) data. Because different regions in the auditory cortex are differentially modulated by attention, the variability in ADI across subjects was thought to be related to the anatomical locations of electrodes. For each subject, and for each frequency band, the pie charts in the graphs 610 and 620 illustrate the proportion of electrodes that were responsive to speech from two (2) anatomical regions: Heschl's gyrus, and the superior temporal gyrus. Electrodes that were responsive to speech, but that were not in either of these locations, are collectively referred to as Other. These Other electrodes were located in various anatomical regions including the middle and inferior temporal gyri, and planum temporale. The single number displayed above each pie chart refers to the total number of electrodes that were responsive to speech for that subject and for the corresponding frequency band.
[0078] To determine which anatomical regions produced the highest ADI, a single electrode analysis was performed, as illustrated in FIG. 6B, which includes graphs 630 and 640 presenting ADI measures per electrode. Similar to FIG. 6A, electrodes are shaded according to anatomical location: Heschi's gyms (FIG; medium dark shade), superior temporal gyrus (STG; dark shade), and Other (light shade). When using HF data, electrodes in STG and Other produced ADIs significantly greater than zero (Wilcoxon signed-rank test; p < 0.001), compared to the use of electrodes in HG (p = 0.02). Using LF data, no region produced ADIs significantly greater than zero (p > 0.05). These results show that electrodes placed in STG are important for successfully decoding attention (an important finding for non-invasive AAD research where source localization methods can be used to target specific brain regions).
[0079] In a further experiment of the implementations described herein, to simulate a dynami c situation where subjects alternated their attention between the two speakers, the data was segmented and concatenated into 60-second chunks, with a subject's attention switching at the beginning of each chunk. FIG. 7 shows the results of the analysis for dynamic switching of attention for an example subject (subject 1) using HF data and a 20 second window size. Each solid vertical line marks a switch in attention. Portion 710a-e of the graph correspond to segments when the subject was attending to Spk2m, and portions 720a-e correspond to segments when the subject was attending to Spklf. Normalized correlation values for each of "the 4 DNNs are plotted at the bottom of the figure. From this analysis, a measure of decoding-accuracy was obtained, which was defined as the percentage of samples in which the correct target speaker produced the highest normalized correlation value. Unlike the ADI method discussed above, this method does not formally take into account false-positives. This is because the sliding window approach inevitably produces false-positives around transitions in attention that are not due to a bias towards a particular speaker, but are instead an inherent property of the methodology.
However, these false-positives will still reduce the decoding-accuracy achievable because the correct target speaker will not produce the highest correlation values in these regions. A decoding-accuracy of 75% was observed for the data displayed in FIG. 7.
[0080] FIG. 8A displays the same results as shown in FIG. 7, but averaged over all sections when the subject was attending to Spk2m (-60s - 0s) and Spklf (0s - 60s). Shaded regions denote standard error. This analysis was used in order to determine a measure of the transition time (how long it takes to detect a switch in attention for a subject). Transition times were
calculated as the time at which the blue (Spk2M) and red (SpklF) lines intersected in the averaged data. This calculation was performed for each subject whose attentiona! focus could be decoded and for all window-sizes (between 5 and 30 seconds: see FIG. 8B). For all subjects, the transition-time increases monotonically with larger window-sizes. Also shown is the decoding- accuracy that was achieved for each window-size (see FIG. 8B; bottom panel). These results suggest that window-sizes between 15 and 20 seconds achieve the best trade-off between speed and accuracy, and that longer window-sizes produce a reduction in decoding-accuracy. Larger window-sizes result in longer transition times, and therefore more false-positives after a switch in attention. FIG. 8C shows an example of the same analysis as in FIG. 8A, but instead uses LF data from subject 2. FIG. 8D illustrates the effect of the window size on transition-time and decoding-accuracy when using the LF data for subjects 1 and 2,
[0081] While the decoding-accuracy score (as presented in FIGS. 8A-D) is an important objective measure of performance, an objective of the experimentations and studies of the implementations described herein was to also enhance the quality of the audio by amplifying the attended speaker relative to the other sound sources. The quality of the output depends on a combination of factors including the accuracy of the DNN speech separation, the neural reconstruction, the attention decoding, and the amplification scheme. The quality of the final output audio was tested using both subjective and objective tests. The output of the system proposed herein produced an objectively cleaner speech signal with a significant increase in the Perceptual Evaluation of Speech Quality (PESQ) score (Wilcoxon signed-rank test, p < 0.001 ; see FIG. 9A showing the PESQ results with the implemented system in OFF state and ON state).
[0082] To subjectively test the preference of users on the quality of the audio output of the systems described herein in a multi-speaker scenario, a psychoacoustic experiment was performed. More particularly, twelve subjects (seven female), aged between 20 and 28 years (mean ± SD, 22 ± 2,5) took part. All subjects reported normal hearing. The stimuli used for this experiment were the same as those used for the neural experiment, i .e. subjects were always presented with a mixture of SpklF and Spk2M. However, the way the stimuli were presented was altered to obtain as much information as possible about the subjects' perception. The experiment was divided into four blocks, each containing fifteen (15) trials. Each trial included a single sentence, Before each block, the subjects were instructed to pay attention to one of the speakers (starting with the male) and to switch attention on each successive block. In order to
test the intelligibility of the speech, after each trial (sentence) the subjects were presented with a transcription of the sentence of the attended speaker with one word missing. Subjects were instructed to type the missing word (the intelligibility task).
[0083] Subjects were also asked to indicate the difficulty they had understanding the attended speaker on a scale from 1 to 5, i.e., very difficult (1), difficult, not difficult, easy, and very easy (5). From these responses, the Mean Opinion Score (MOS) was calculated. This allowed obtaining both an objective measure of intelligibility and a subjective measure of listening effort (MOS). For half of the experiment, both speakers were presented at the same RMS power. For the other half, an attempt was made to amplify the attended speaker. Block order was counterbalanced across subjects, in total, the experiment lasted approximately 15 minutes, during which subjects were presented with 4 minutes and 11 seconds of audio. FIG. 9B is a graph showing the MOS scores when the system was off and on, and illustrates that MOS measures were significantly improved when the system was on. Particularly, all but one subject reported a larger MOS when the system was on, with a median MOS of 3.87 (25iJl percentile, 3.73; 75ib percentile, 4.12) versus a median MOS of 3.5 (25th percentile, 2.9; 75th percentile, 3.8) when the system was off. The one subject who reported no reduction in listening effort had almost identical MOS scores for both system on (3.7) and system off (3.8). The majority of subjects also reported a preference for the segments where the system was turned on (9 out of 12), and a majority reported a preference for using the system once they were informed of its nature (10 out of 12), In the missing word task, there was no significant difference in the numbers of words that were correctly reported when the system was on versus off. This lack of improved intelligibility is a well-known phenomenon in speech enhancement research where noise suppression does not typically improve intelligibly scores, even though the listening effort is reduced,
[0084] As part of the psychoacoustic experimentation conducted herein, real neural data was used to demonstrate how the overall system could be implemented. Specifically, the neural data from one of the subjects (subject 2) was used. To dynamically track the attentional focus of the subject, a strategy similar to the artificial switching of attention discussed earlier was
implemented, i.e., use a sliding window approach, attempting to decode the attention of the subject every second, LF data was used, as opposed to the HF data, to be as comparable as possible with noninvasive technologies which typically only have access to low-frequency neural activity. Additionally, a window size of 20 seconds was used in order to be consistent with the
decoding strategies discussed herein. Whenever it was possible to correctly classify the attended speaker from the neural data for that subject, the output from the correct DNN was added to the mixture. However, if a mistake was made, and the subject's attentional focus was misclassified, the output from whichever DNN produced the largest normalized correlation was presented. The DNN output was added at a level of +12dB relative to the mixture, hi addition to obtaining measures of intelligibility and listening effort, a determination of the participants' overall preference was made. Subjects were informed before the experiment that they would have to report which half of the experiment requi red less effort to understand the attended speaker, and were reminded half way through. The experimentation also asked the users as to whether they would prefer to use the system if they knew that it was not perfect.
[0085] Thus, as discussed herein, an end-to-end system that incorporates a single channel automatic speech-separation process into the auditory attention-decoding (AAD) platform was described. The approaches discussed herein alleviate the spatial separation requirements of multi-channel approaches, but can also be used in tandem with beamforming methods for optimal source separation. In addition to successfully identifying the attended speaker, the implementations described herein also amplify that speaker, resulting in an improvement in the subjective quality of the listening experience. Combined with the latest developments in AAD research, the systems described herein can be used to realize hearing aid devices that can automatically and dynamically track a user's direction of attention, and amplify an attended speaker (and/or attenuate non-attended speaker(s)). In some embodiments, neurological signals can be obtained through invasive ECoG recordings from neurological patients, which allows examining the upper bound of decoding speed and accuracy, determining the varying contributions of different auditory cortical areas, and separately examining attention decoding using low (l-8Hz) and high (70-150Hz) frequency neural data. In some embodiments, neural data may be obtained also through non-invasive neural recordings.
[0086] While the subjects in the experiments conducted for the implementations described herein were only presented with mixtures of two speakers (SpklF and Spk2M), the
implementations can be extended to a more general case of multiple speakers. Each DNN of the implementation may therefore be able to separate a trained target speaker from other unseen speakers.
[0087] The proposed system can switch to default operation when no target speaker is detected. Generally, a new speaker can be easily added to the system, requiring only a small amount, of clean speech from the new target speaker (e.g., -20 minutes). Because the reconstruction process from neural data may be achieved using single speaker data training, adding a new speaker does not involve recording new neural responses for that speaker. A practical limitation for all processes intended for hearing aids (or other hearing augmentation devices) is that hardware constraints could limit the number of DNNs that could be housed inside a portable device. However, modem hearing aids are able to perform off-board computing by interfacing with a ceil phone, and specialized hardware is also becoming available for low-power neural network implementations. Another consideration is the fact that DNNs rely heavily on the data used to train them, Therefore, additional training may be required to separate speakers under different environmental conditions. Also, because people tend to involuntarily speak louder in noisy situations, which affects acoustic features such as pitch, rate and syllable duration, this would also need to be taken into account during training of the DNNs.
[0088] It was determined that the reconstructed spectrograms in data had a high correlation with the raw mixture. This is a problem that needs to be addressed, because a DNN is likely to output the mixture when its designated speaker is not present. The reason why the reconstructed spectrograms had such a high correlation with the mixture can be understood by looking at the single-electrode analysis. It was determined that many speech responsive electrodes are not modulated by attention, and instead encode both the attended and unattended speakers.
Therefore, training decoders using all speech-responsive electrodes will lead to reconstructions that are similar to the mixture. The analysis also revealed that the location of electrodes in the brain played an important role in decoding attention. This result is likely due to the variable degree of attentionai modulation in various parts of the auditory system. There was also a dichotomy between the high and low frequency data, where the accuracy of decoding was not. always related, possibly due to the different neural generators of these two frequency bands.
[0089] In a real-world situation, it is likely that users would want to dynamically switch their attention between multiple speakers as a conversation progresses. Although the subjects in the experiments conducted for the implementations described herein alternated their attention between two speakers, they did not do so in a dynamic fashion, Rather, there was a substantial break between each block of the experiment. When simulating switching of attention, the
window size used for estimation has an effect on both decoding-accuracy and transition-time (the time it takes to detect a switch in attention). The results of the experiments indicated that there is an optimal window-size for decoding-accuracy: shorter window sizes produce r-values that are too noisy, and longer window sizes prohibit the rapid detection of switches in attention. This problem is particularly important when using neural signals with a lower signal to noise ratio (such as around the ear, or in ear EEG). It is possible that more elaborate decoding processes can be used to speed up decoding and provide a better trade-off between decoding-accuracy and transition-time.
[0090] One important requirement of speech enhancement techniques is to ensure that the resulting speech is not distorted or corrupted, as users tend to prefer no enhancement over an amplified but distorted signal. In the experiments conducted, the automatically separated target speaker was amplified by +12dB relative to the mixture. This level has been shown to significantly increase the intelligibility of an attended speaker in a two-talker scenario (from -88% to 98%). Importantly, an unattended speaker should still be audible so that users can switch their attention should they choose to do so. It is still possible to understand speakers when they are attenuated by I 2dB, although intelligibility drops to ~ 78%. These parameters need to be further optimized when this type of system is tested in hearing impaired listeners and in closed-loop setups, where the decoding processes and a subject's brain can co-adapt to converge to a suitable solution. It was determined that there is a noticeable increase in the mean opinion score (MOS) when the implementations described herein were used, and almost all subjects reported that they would prefer to have the system turned on, a finding that supports the implementations' potential as a useful and effective way to identify and amplify an attended speaker.
Deep Attractor Network for Signal Separation
[0091] As noted, in some embodiments, a deep learning framework, also referred to as the attractor network, is implemented to solve the source separation problem. The term attractor is motivated by the well-studied perceptual effects in human speech perception which suggest that the brain circuits create perceptual attractors (magnets) that warp the stimulus space so as to draw the sound that is closest to it, a phenomenon that is called Perceptual Magnet Effect. The
proposed model works on the same principle by forming a reference point (attractor) for each source in the embedding space which draws all the T-F bins toward itself. Using the similarity between the embedded points and each attractor, a mask is estimated for each sources in the mixture. Since the mask is directly related to the attractor point, the proposed framework can potentially be extended to arbitrary number of sources without the permutation problem.
Moreover, the mask learning allows a very efficient end-to-end training scheme and highly reduces the computation complexity compared with the deep-clustering (DC) approach and the permutation invariant training (PIT) approach. In deep clustering, a network is trained to generate discriminative embedding for each time-frequency (T-F) bin with points belonging to the same source forced to be closer to each other. DC is generally able to solve both permutation and output dimension problem to produce the state of the art separation performance. The main drawback of DC is its relative inefficiency to perform end-to-end mapping, because the objective function is the affinity between the sources in the embedded space and not the separated signals themselves. Minimizing the separation error is done with an unfolding clustering system and a second network, which is trained iteratively and stage-by-stage to ensure convergence. The PIT algorithm solves the permutation problem by pooling over all possible permutations for N mixing sources (jV! permutations), and using the permutation with lowest error to update the network. The PIT approach was shown to have comparable performance as DC. However, PIT approach suffers the output dimension mismatch problem because it assumes a fixed number of sources. PIT also suffers from its computation efficiency, where the prediction window has to be much shorter than context window due to the inconsistency of the permutation both across and within sample segments.
[0092] With reference to FIG. 10, a diagram of a proposed system architecture 1000 to implement a deep-attractor network approach is provided. As shown, the system 1000 receives as input a mixture signal 1002 (e.g., a multi-speaker combined signal) via an input sensor (e.g., a single or multiple microphones coupled to a device such as a hearing devices, or some other hearing augmentation device). The received input mixture signal 1002 is provided to a deep learning neural network 1010, which may be implemented, for example, as an LSTM layer. In some embodiments, the neural network is trained to map the mixture sound X to a k dimensional embedding space (represented by the box 1020), such that it minimizes an objective function such as the following:
where S is the clean spectrogram (frequency F x time 7) of C sources, the mixture
objective functions may also be used.
[0093] The mask is estimated in the K dimensional embedding space of each T-F bin,
represented by
where
are the attractors (represented in box 1030) for the C sources in the embedding space, learned during training, which are defined as
where is the source membership function for each T-F bin, i.e., if source c
has the highest energy at time / and frequency/ compare to the other sources.
[0094] The objective function in Equation 7 includes of three parts. During training, an embedding Vis first computed through a forward pass of the neural network for each given mixture. Then an attractor vector is estimated for each source using Equation 9, This can be done in several ways. The most straightforward method for attractor generation is to find the source centroid, as defined in Equation 9.
[0095] Next, a reconstruction mask is estimated (at box 1040) for each source by finding the similarity of each T-F bin in the embedding space to each of the attractor vectors A, where the similarity metric is defined in Equation 8. This particular metric uses the inner product followed by a sigmoid function which monotoiiicaily scales the masks between [0, 1], Intuitively, if an embedding of a T-F bin is closer to one attractor, then it means that it belongs to that source, and the resulting mask for that source will produce larger values for that T-F bin. Since the source separation masks for each TF bin should add up to one, particularly in difficult conditions, the sigmoid function (of Equation 8) can be replaced with softmax function:
[0096] Finally, a standard L2 reconstruction error is used to generate the gradient, as shown in Equation 7. Therefore, the error for each source reflects the difference between the masked signal and the clean reference, forcing the network to optimize the global reconstruction error for better separation. The proposed net is referred to as deep attractor network (DANet).
[0097] As noted, the DANet approach has several advantages. Firstly, DANet removes the stepwise pre-training process (required, for example, in the DC methodology) to allow end-to- end training. Another advantage of the DANet approach arises from flexibility in the number of sources in the mixture.
[0098] Attractor points can be estimated using various processes other than the average-based approach. One possibility is to use weighted average. Since the attractors represent the sources' center of gravity, only the embeddings of the most salient T-F bins may be included, which leads to more robust estimation. This strategy can be investigated by using an amplitude threshold in the estimation of the attractor. Alternatively, a neural network model may also be used to pick the representative embedding for each source, an idea which shares similarities with encoder- decoder attention networks. During test time, because the true assignment Y is unknown, two strategies can be applied to form the attractor points. The first is similar to the strategy used in DC, where the centers are found using post K-means process. The second strategy is based on the observation that the location of the attractors in the embedding space is relatively stable. This observation is shown in FIGS. 11 A-B, which includes, at FIG. 11 A, graphs showing an example of a mixture, the difference between two speakers in the mixture, and the separated spectrograms of the two speakers using a DANet implementation, and, at FIG. 1 IB, a graph showing locations of T-F bins in the embedding space, where each pair of dots corresponds to the attractor found for the two speakers in a given mixture.
[0099] The objective function of DC is shown in Equation 11, where Y is the indicator function which is equivalent to a binary mask, and V is the learned embedding:
[00100] Since 7 is orthogonal and constant for each mixture, by multiplying Y1 and a normalizer to both terms, an objective function can be derived that is a special case
[00101] In Equation can be viewed as an averaging operation, where the
embeddings are summed according to the label, and the resulted center is multiplied with the embedding matrix Fto measure the similarity between each embedding and the center, and compared with the ground truth binary mask. When the learned Fis optimum, i.e., VV! YY1, Equations 11 and 12 are equivalent. On the other hand, when the attractor vectors are considered as free parameters in the network, DANet reduces to a classification network, and Equation 7 becomes a fully-connected layer. In this case, PIT may become necessaiy because the mask has no information about the source and the problem of fixed output dimension arises. In contrast, the freedom of the network to form attractor points during the training allows the system to use the affinity between samples where no constraint is on the number of patterns found, therefore allowing the network to be independent of the nuraber of sources. The flexibility of the network in choosing the attractor points is helpful even in two-source separation problems, because the two sources may have very different structures. As can be seen in FIG. 12, providing a graph 1200 presenting the locations of attractor points in the embedding space, the proposed implementations trained in speaker separation tasks have ended up finding 2 attractor pairs (4 points in the embedding space). More particularly, each point in FIG. 12 corresponds to one of 10,000 mixture sounds, visualized using the first three principal, components. Two distinct attractor pairs (denoted as A l and A2) are visible.
[00102] With reference now to FIG. 13, a flowchart of an example deep attractor signal separation procedure 1300 is shown. The procedure 1300 includes acquiring 1310 a mixture speech signal from multiple sources, the mixture signal defining a time-frequency embedding space. As discussed herein, the mixture signal may be the combined sounds signal of the procedure 300, which includes multiple sounds signals from multiple sounds sources. As also notes, the mixture speech signal may be obtained using a single audio sensor, such as a microphone, or multiple audio sensors. Having obtained the mixture speech signal, the procedure 1300 further includes determining 1320 respective reference points for each of the
multiple sources, with the reference points representing locations of the sources in the embedding space for the mixture speech signal. The reference points may include attractor points in the embedding space. In some embodiments, determining the respective reference points may include determining the respective reference points using a deep neural network,
[00103] At 1330, the procedure 1300 further includes deriving masks for the determined reference points, and extracting 1340 at least one of the multiple sources using at least one of the derived masks. Deriving the masks may include computing similarity between embedded time- frequency points within the embedding space and the determined respective reference points.
[00104] As noted, the deep attractor network approach for speech separati on may be used in many different application. For example, the DANet approach may be used to separate multi- speaker speech signals to implement attentional selection processing (i.e., enhance the speech signals of a speaker that the listener is determined to be listening to or focused on). In another example, the DANet approach may also be used in surveillance applications (to separate and record speech from a combined sound signal corresponding to different speakers), for virtual reality applications (reconstruct speaking objects from a pre-recorded source, and render the separated speech signals into a virtual audio, or audio-visual, scene), etc.
[00105] The deep attractor network implementations described herein were tested to evaluate their performance and efficacy. To that end, a system was realized for which a 30 hour training set and a 10 hour validation set, generated by randomly selecting utterances from different speakers in the Wall Street Journal (WSJO) training data set si_tr_s was used. The data was mixed with various signal-to-noise ratios (SNR) randomly chosen between 0 dB and 10 dB. A five (5) hour evaluation set was generated similarly as above, using utterances from sixteen (16) unseen speakers from si dt 05 and si et 05 data in the WSJO dataset. Additionally, a three speaker mixture dataset was constructed for a three speaker separation evaluation from the same WSJ set, which had 30 hour training data, 10 hours validation data, and 5 hours testing data, mixed with SNR at -5~ 5 dB. Each three speaker mixture was ensured to include both female and male speakers. All data was re-sampled to 8 kHz to reduce computational and memory costs. The log spectral magnitude was served as an input feature, computed using short-time Fourier transform (STFT) with 32 ms window length, 8 ms hop size, and the square root of harming window.
[00106] The network contained four (4) bi-directional LSTM layers with 600 hidden units in each layer. The embedding dimension was set to 20, resulting in a fully-connected feedforward layer of 2580 hidden units (20 x 129) after the BLSTM layers. The input features were split into non-overlapping chunks of 100-frame length as the input to the network. An RMSprop procedure was used for training, with an exponential learning rate decaying strategy, where the learning rate started at 10-4 and ended at 3 x 10-6, The total number of epochs was set to be 150, and a cost function was used in Equation 7 on the validation set for early stopping. The criteria for early stopping is no decrease in the loss function on validation set for 10 epochs. A Deep Clustering (DC) network was constructed with the same configuration, which was used as the baseline. The results were reported in terms of signal-to-distortion ratio (SDR, which was defined as scale-invariant SNR here), signal-to-artifacts ratio (SAR), and signal-to-interference ratio (SIR). The results are shown below in Table 1, which provides evaluation metrics for networks with different configurations.
[00107] In Table 1 above, the percentage indicated represents the salient weigh threshold used during training. As indicated by the results of Table 1, although the plain DANet already outperforms the DC baseline, adding a simple threshold on T-F samples included in the formation of the attractor yields further improved performance, presumably due to the increased
influence of salient segments. On the other hand, the performance boost suggests that better attractor formation procedures can be utilized to estimate the attractors, such as joint optimizing of the network parameters. Moreover, by applying curriculum training strategy, in which the network continues to be trained with 400-frame length input, DANet achieves the best overall performance.
[00108] In the last experiment presented in Table 1, a fixed pair of attention vector collected from the training data was used, corresponding to the Al pair of FIG. 12, This pre-set attractor is able to generalize well to the unseen mixtures and produces high quality separation, but corresponds to slightly worse results than the best model . Compared with K -means, this has the advantage that it can be easily implemented in real-time using a frame-by-frame pipeline. Based on this observation, when more attractors are required (e.g. in more complex tasks), a collection of attractor codebook and a simple classifier could be implemented for real-time processing,
[00109] Next, in a three-speaker separation experiment, whose results are shown in Table 2 below, the proposed system significantly outperforms the deep clustering baseline. This result may be expected since deep clustering was trained to estimate binary mask, while the deep attractor network focuses on the signal reconstmction. When the mixture is relatively simple, the binary mask could generate high quality separation. However, for more complex mixtures, or when one source is significantly louder than the other, the binary mask usually leads to large bias to the loudest source, and thus result in unsatisfactory separation for the weaker source. Note that in the three speaker experiment, the network was trained using softmax objective as shown in Equation 10.
[00110] Thus, described herein are implementations of a neural network framework referred to as a deep attractor network for general source separation problem. The network forms attractor points in a high-dimensional embedding space of the signal, and the similarity between attractors and time-frequency embeddings are then converted into a soft separation mask. The proposed framework can perform end-to-end, real-time separation, and can work on different number of mixing sources. Evaluations and experimentations performed for the deep attractor network implementations for multi-speaker separation tasks confirmed its efficacy and potential for extension to general source separation problems.
Other Embodiments
[00111] Performing the various procedures, processes, and operations described herein may be facilitated by a controller system (e.g., a processor-based controller system, a state machine, etc.) Particularly, at least some of the various devices/systems (e.g., hearing devices, or other types of hearing augmentation devices) described herein may be implemented, at least in part, using one or more controllers such as the one shown in FIG, 14, illustrating a generic computing or controller system 1400, The computing system 1400 includes a computing-based device 1410 such as a personal computer, a specialized computing device, a controller circuit (implemented on a chip), and so forth, that typically includes a central controller 1412 (which may a programmable processor, such as a CPU). In addition to the central controller 1412, the system includes main memory, cache memory and bus interface circuits (not shown in FIG. 14). The computing / controller-based device 1410 may include a mass storage element 1414, such as a hard drive or flash drive associated with the computer system. The computing / controller system 1400 may further include a keyboard 1416, or keypad, or some other user input interface, and a monitor 1420, e.g., an LCD (liquid crystal display) monitor, that may be placed where a user can access them. The computing system 1400 may be incorporated within a system such as the various devices described herein.
[00112] The computing-based device 1410 is configured to facilitate, for example, the implementation of operations for neural decoding of attention al selection in multi -speaker environments and/or for deep attractor networks. The storage device 1414 may thus include a computer program product that when executed on the computing / controller-based device 1410
causes the computing-based device to perform operations to facilitate the implementation of procedures and operations described herein. The computing / controller-based device may further include peripheral devices to enable input/output functionality. Such peripheral devices may include, for example, a CD-ROM drive and/or flash drive (e.g., a removable flash drive), or a network connection (e.g., implemented using a USB port and/or a wireless transceiver), for downloading related content to the connected system. Such peripheral devices may also be used for downloading software containing computer instaictions to allow general operation of the respective system/device. Alternatively and/or additionally, in some embodiments, special purpose logic circuitry, e.g., an FPGA (field programmable gate array), an ASIC
(application-specific integrated circuit), a DSP processor, etc., may be used in the
implementation of the system 1400. Other modules that may be included with the computing- based device 1410 are speakers, a sound card, a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computing system 1400. The computing / controller- based device 1410 may include an operating system, e.g., Windows XP® Microsoft Corporation operating system, Ubuntu operating system, etc,
[00113] Computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and may be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language. As used herein, the term "'machine-readable medium'" refers to any non-transitory computer program product, apparatus and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a non-transitory machine-readable medium that receives machine instructions as a machine-readable signal.
[00114] In some embodiments, any suitable computer readable media can be used for storing instructions for performing the processes / operations / procedures described herein. For example, in some embodiments computer readable media can be transitory or non-transitory. For example, non-transitory computer readable media can include media such as magnetic media (such as hard disks, floppy disks, etc.), optical media (such as compact discs, digital video discs, Blu-ray discs, etc.), semiconductor media (such as flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read only Memory (EEPROM), etc.), any suitable media that is not fleeting or not devoid of any semblance of permanence
during transmission, and/or any suitable tangible media. As another example, transitory computer readable media can include signals on networks, in wires, conductors, optical fibers, circuits, any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and/or any suitable intangible media.
[00115] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly or conventionally understood. As used herein, the articles "a" and "an" refer to one or to more than one (i .e., to at least one) of the grammatical object of the article. By way of example, "an element" means one element or more than one element. "About" and/or "approximately" as used herein when referring to a measurable value such as an amount, a temporal duration, and the like, encompasses variations of ±20% or ±10%, ±5%, or +0.1% from the specified value, as such variations are appropriate in the context of the systems, devices, circuits, methods, and other implementations described herein. "Substantially" as used herein when referring to a measurable value such as an amount, a temporal duration, a physical attribute (such as frequency), and the like, also encompasses variations of ±20% or ±10%, ±5%, or +0.1% from the specified value, as such variations are appropriate in the context of the systems, devices, circuits, methods, and other implementations described herein.
[00116] As used herein, including in the claims, "or" as used in a list of items prefaced by "at least one of or "one or more of indicates a disjunctive list such that, for example, a list of "at least one of A, B, or C" means A or B or C or AB or AC or BC or ABC (i.e., A and B and C), or combinations with more than one feature (e.g., AA, AAB, ABBC, etc.). Also, as used herein, unless otherwise stated, a statement that a function or operation is "based on" an item or condition means that the function or operation is based on the stated item or condition and may be based on one or more items and/or conditions in addition to the stated item or condition.
[00117] Although particular embodiments have been disclosed herein in detail, this has been done by way of example for purposes of illustration only, and is not intended to be limiting with respect to the scope of the appended claims, which follow. Some other aspects, advantages, and modifications are considered to be within the scope of the claims provided below. The claims presented are representative of at least some of the embodiments and features disclosed herein. Other unclaimed embodiments and features are also contemplated.
Claims
1. A method comprising:
obtaining, by a device, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located;
applying, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources; obtaining, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to;
selecting one of the plurality of separated signals based on the obtained neural signals for the person; and
processing the selected one of the plurality of separated signals.
2. The method of claim 1 , further comprising:
generating a sound spectrogram from the combined sound signal; and
applying the neural-network-based speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
3. The method of claim 2, wherein selecting the one of the plurality of separated sound signals based on the obtained neural signals for the person comprises:
generating an attended speaker spectrogram based on the neural signals for the person; comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms; and
transforming the selected one of the multiple resultant speaker spectrograms into an acoustic signal.
4. The method of claim 3, wherein comparing the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms comprises:
comparing the attended speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
5. The method of claim 1, wherein obtaining the neural signals for the person comprises: obtaining one or more of: electrocorticography (ECoG) signals for the person, neural measurements via a non-invasive scalp, or in-ear EEG recordings.
6. The method of claim 1, wherein applying the neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources comprises:
providing the combined sound signal from the multiple sound sources to a deep neural network (DNN) configured to identify individual sound sources from the combined sound signal.
7. The method of claim 1, wherein processing the selected one of the plurality of separated sound signals comprises one or more of: amplifying the selected one of the plurality of separated signals, or attenuating at least one non-selected signal from the plurality of separated signals.
8. The method of claim 1, wherein obtaining the combined sound signal for the multiple sound sources comprises:
receiving the combined sound signal for the multiple sound sources at a single microphone coupled to the device.
9. The method of claim 1, wherein the combined sound signal defines a time- frequency embedding space, and wherein applying the neural-network-based speech-separation processing to the combined sound signal comprises:
determining respective reference points for each of the multiple sound sources, with the reference points representing locations of the sound sources in the embedding space for the combined sound signal;
deriving masks for the determined reference points; and
extracting at least one of the multiple sound sources using at least one of the derived masks.
10. the method of claim 9, wherein determining the respective reference points comprises:
determining the respective reference points using a deep neural network.
11. The method of claim 9, wherein deriving the masks comprises:
computing similarity between embedded time-frequency points within the embedding space and the determined respective reference points.
12. The method of claim 9, wherein the reference points comprise attractor points in the embedding space.
13. A system comprising:
at least one microphone to obtain a combined sound signal for signals combined from multiple sound sources in an area in which a person is located;
one or more neural sensors to obtain neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to; and
a controller coupled to the at least one microphone and the one or more neural sensors, the controller configured to:
apply neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources;
select one of the plurality of separated signals based on the obtained neural signals for the person; and
process the selected one of the plurality of separated signals.
14. The system of claim 13, wherein the controller is further configured to:
generate a sound spectrogram from the combined sound signal; and
apply the speech-separation processing to the generated spectrogram to derive multiple resultant speaker spectrograms.
15. The system of claim 14, wherein the controller configured to select the one of the plurality of separated signals based on the obtained neural signals for the person is configured to: generate an attended speaker spectrogram based on the neural signals for the person;
compare the attended speaker spectrogram to the derived multiple resultant speaker spectrograms to select one of the multiple resultant speaker spectrograms; and
transform the selected one of the multiple resultant speaker spectrograms into an acoustic signal.
16. The system of claim 15, wherein the controller configured to compare the attended speaker spectrogram to the derived multiple resultant speaker spectrograms is configured to: compare the attended-speaker spectrogram to the derived multiple resultant speaker spectrograms using normalized correlation analysis.
17. The system of claim 13, wherein the one or more neural sensors comprise one or more of: an electrocorticography (ECoG) sensor to measure ECoG signals for the person, a noninvasive scalp to obtain neural measurements, or in-ear EEG sensor to obtain in-ear EEG recordings.
18. The system of claim 13, wherein the combined sound signal defines a time- frequency embedding space, and wherein the controller configured to apply the neural-network-based speech-separation processing to the combined sound signal is configured to:
determine respective reference points for each of the multiple sound sources, with the reference points representing locations of the sound sources in the embedding space for the combined sound signal;
derive masks for the determined reference points; and
extract at least one of the multiple sound sources using at least one of the derived masks.
19. the system of claim 18, wherein the controller configured to determine the respective reference points is configured to:
determine the respective reference points using a deep neural network.
20. The system of claim 18, wherein the controller configured to derive the masks is configured to:
compute similarity between embedded time- frequency points within the embedding space and the determined respective reference points.
21. A non-transitory computer readable media programmed with instructions, executable on a processor, to:
obtain, by a device comprising the processor, a combined sound signal for signals combined from multiple sound sources in an area in which a person is located;
apply, by the device, neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources;
obtain, by the device, neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to;
select one of the plurality of separated signals based on the obtained neural signals for the person; and
process the selected one of the plurality of separated signals.
22. An apparatus comprising:
means for obtaining a combined sound signal for signals combined from multiple sound sources in an area in which a person is located;
means for applying neural-network-based speech-separation processing to the combined sound signal from the multiple sound sources to derive a plurality of separated signals that each contains signals corresponding to different groups of the multiple sound sources;
means for obtaining neural signals for the person, the neural signals being indicative of one or more of the multiple sound sources the person is attentive to;
means for selecting one of the plurality of separated signals based on the obtained neural signals for the person; and
means for processing the selected one of the plurality of separated signals.
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP17813908.5A EP3469584B1 (en) | 2016-06-14 | 2017-06-13 | Neural decoding of attentional selection in multi-speaker environments |
US16/169,194 US11373672B2 (en) | 2016-06-14 | 2018-10-24 | Systems and methods for speech separation and neural decoding of attentional selection in multi-speaker environments |
US17/826,474 US11961533B2 (en) | 2016-06-14 | 2022-05-27 | Systems and methods for speech separation and neural decoding of attentional selection in multi-speaker environments |
US18/228,239 US20240013800A1 (en) | 2016-06-14 | 2023-07-31 | Systems and methods for speech separation and neural decoding of attentional selection in multi-speaker environments |
Applications Claiming Priority (6)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201662349976P | 2016-06-14 | 2016-06-14 | |
US62/349,976 | 2016-06-14 | ||
US201662429549P | 2016-12-02 | 2016-12-02 | |
US62/429,549 | 2016-12-02 | ||
US201762480005P | 2017-03-31 | 2017-03-31 | |
US62/480,005 | 2017-03-31 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US16/169,194 Continuation-In-Part US11373672B2 (en) | 2016-06-14 | 2018-10-24 | Systems and methods for speech separation and neural decoding of attentional selection in multi-speaker environments |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2017218492A1 true WO2017218492A1 (en) | 2017-12-21 |
Family
ID=60663355
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2017/037186 WO2017218492A1 (en) | 2016-06-14 | 2017-06-13 | Neural decoding of attentional selection in multi-speaker environments |
Country Status (2)
Country | Link |
---|---|
EP (1) | EP3469584B1 (en) |
WO (1) | WO2017218492A1 (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108198569A (en) * | 2017-12-28 | 2018-06-22 | 北京搜狗科技发展有限公司 | A kind of audio-frequency processing method, device, equipment and readable storage medium storing program for executing |
CN108847238A (en) * | 2018-08-06 | 2018-11-20 | 东北大学 | A kind of new services robot voice recognition methods |
CN109124625A (en) * | 2018-09-04 | 2019-01-04 | 大连理工大学 | A kind of driver fatigue state horizontal mipmap method |
WO2019199554A1 (en) * | 2018-04-11 | 2019-10-17 | Microsoft Technology Licensing, Llc | Multi-microphone speech separation |
WO2019223457A1 (en) * | 2018-05-24 | 2019-11-28 | 腾讯科技(深圳)有限公司 | Mixed speech recognition method and apparatus, and computer readable storage medium |
CN110675891A (en) * | 2019-09-25 | 2020-01-10 | 电子科技大学 | Voice separation method and module based on multilayer attention mechanism |
US20200237296A1 (en) * | 2017-03-02 | 2020-07-30 | Cornell University | A sensory evoked diagnostic for the assessment of cognitive brain function |
CN112528896A (en) * | 2020-12-17 | 2021-03-19 | 长沙理工大学 | SAR image-oriented automatic airplane target detection method and system |
CN113113000A (en) * | 2021-04-06 | 2021-07-13 | 重庆邮电大学 | Lightweight speech recognition method based on adaptive mask and grouping linear transformation |
CN113314136A (en) * | 2021-05-27 | 2021-08-27 | 西安电子科技大学 | Voice optimization method based on directional noise reduction and dry sound extraction technology |
US11170785B2 (en) | 2016-05-19 | 2021-11-09 | Microsoft Technology Licensing, Llc | Permutation invariant training for talker-independent multi-talker speech separation |
CN113647119A (en) * | 2019-01-25 | 2021-11-12 | 索诺瓦有限公司 | Signal processing apparatus, system and method for processing audio signals |
CN113812173A (en) * | 2019-05-09 | 2021-12-17 | 索诺瓦有限公司 | Hearing device system and method for processing audio signals |
WO2022076404A1 (en) * | 2020-10-05 | 2022-04-14 | The Trustees Of Columbia University In The City Of New York | Systems and methods for brain-informed speech separation |
US11335332B2 (en) | 2019-09-03 | 2022-05-17 | Stmicroelectronics S.R.L. | Trigger to keyword spotting system (KWS) |
US11456007B2 (en) | 2019-01-11 | 2022-09-27 | Samsung Electronics Co., Ltd | End-to-end multi-task denoising for joint signal distortion ratio (SDR) and perceptual evaluation of speech quality (PESQ) optimization |
CN115153563A (en) * | 2022-05-16 | 2022-10-11 | 天津大学 | Mandarin auditory attention decoding method and device based on EEG |
US11495215B1 (en) * | 2019-12-11 | 2022-11-08 | Amazon Technologies, Inc. | Deep multi-channel acoustic modeling using frequency aligned network |
US11630513B2 (en) | 2018-12-20 | 2023-04-18 | Massachusetts Institute Of Technology | End-to-end deep neural network for auditory attention decoding |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20010018652A1 (en) * | 1996-01-15 | 2001-08-30 | Stephen Mclaughlin | Waveform synthesis |
US20070083365A1 (en) * | 2005-10-06 | 2007-04-12 | Dts, Inc. | Neural network classifier for separating audio sources from a monophonic audio signal |
US20090116652A1 (en) * | 2007-11-01 | 2009-05-07 | Nokia Corporation | Focusing on a Portion of an Audio Scene for an Audio Signal |
US20090279715A1 (en) * | 2007-10-12 | 2009-11-12 | Samsung Electronics Co., Ltd. | Method, medium, and apparatus for extracting target sound from mixed sound |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5150323A (en) * | 1989-08-11 | 1992-09-22 | Hughes Aircraft Company | Adaptive network for in-band signal separation |
US9390712B2 (en) * | 2014-03-24 | 2016-07-12 | Microsoft Technology Licensing, Llc. | Mixed speech recognition |
-
2017
- 2017-06-13 EP EP17813908.5A patent/EP3469584B1/en active Active
- 2017-06-13 WO PCT/US2017/037186 patent/WO2017218492A1/en unknown
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20010018652A1 (en) * | 1996-01-15 | 2001-08-30 | Stephen Mclaughlin | Waveform synthesis |
US20070083365A1 (en) * | 2005-10-06 | 2007-04-12 | Dts, Inc. | Neural network classifier for separating audio sources from a monophonic audio signal |
US20090279715A1 (en) * | 2007-10-12 | 2009-11-12 | Samsung Electronics Co., Ltd. | Method, medium, and apparatus for extracting target sound from mixed sound |
US20090116652A1 (en) * | 2007-11-01 | 2009-05-07 | Nokia Corporation | Focusing on a Portion of an Audio Scene for an Audio Signal |
Non-Patent Citations (2)
Title |
---|
EYNDHOVEN ET AL.: "EEG-informed attended speaker extraction from recorded speech mixtures with application in neuro-steered hearing prostheses", IEEE TRANS. BIOMEDICAL ENGINEERING, vol. 64, no. 5, 18 February 2016 (2016-02-18), pages 1045 - 1056, XP011646202, Retrieved from the Internet <URL:https://arxlv.org/abs/1602.05702> [retrieved on 20170825] * |
See also references of EP3469584A4 * |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11170785B2 (en) | 2016-05-19 | 2021-11-09 | Microsoft Technology Licensing, Llc | Permutation invariant training for talker-independent multi-talker speech separation |
US20200237296A1 (en) * | 2017-03-02 | 2020-07-30 | Cornell University | A sensory evoked diagnostic for the assessment of cognitive brain function |
US11759146B2 (en) * | 2017-03-02 | 2023-09-19 | Cornell University | Sensory evoked diagnostic for the assessment of cognitive brain function |
CN108198569B (en) * | 2017-12-28 | 2021-07-16 | 北京搜狗科技发展有限公司 | Audio processing method, device and equipment and readable storage medium |
CN108198569A (en) * | 2017-12-28 | 2018-06-22 | 北京搜狗科技发展有限公司 | A kind of audio-frequency processing method, device, equipment and readable storage medium storing program for executing |
US10957337B2 (en) | 2018-04-11 | 2021-03-23 | Microsoft Technology Licensing, Llc | Multi-microphone speech separation |
WO2019199554A1 (en) * | 2018-04-11 | 2019-10-17 | Microsoft Technology Licensing, Llc | Multi-microphone speech separation |
WO2019223457A1 (en) * | 2018-05-24 | 2019-11-28 | 腾讯科技(深圳)有限公司 | Mixed speech recognition method and apparatus, and computer readable storage medium |
US11996091B2 (en) | 2018-05-24 | 2024-05-28 | Tencent Technology (Shenzhen) Company Limited | Mixed speech recognition method and apparatus, and computer-readable storage medium |
CN111445905A (en) * | 2018-05-24 | 2020-07-24 | 腾讯科技(深圳)有限公司 | Hybrid speech recognition network training method, hybrid speech recognition device and storage medium |
EP3806089B1 (en) * | 2018-05-24 | 2024-06-19 | Tencent Technology (Shenzhen) Company Limited | Mixed speech recognition method and apparatus, and computer readable storage medium |
CN111445905B (en) * | 2018-05-24 | 2023-08-08 | 腾讯科技(深圳)有限公司 | Mixed voice recognition network training method, mixed voice recognition method, device and storage medium |
CN108847238A (en) * | 2018-08-06 | 2018-11-20 | 东北大学 | A kind of new services robot voice recognition methods |
CN109124625A (en) * | 2018-09-04 | 2019-01-04 | 大连理工大学 | A kind of driver fatigue state horizontal mipmap method |
CN109124625B (en) * | 2018-09-04 | 2021-07-20 | 大连理工大学 | Driver fatigue state level grading method |
US11630513B2 (en) | 2018-12-20 | 2023-04-18 | Massachusetts Institute Of Technology | End-to-end deep neural network for auditory attention decoding |
US11456007B2 (en) | 2019-01-11 | 2022-09-27 | Samsung Electronics Co., Ltd | End-to-end multi-task denoising for joint signal distortion ratio (SDR) and perceptual evaluation of speech quality (PESQ) optimization |
CN113647119A (en) * | 2019-01-25 | 2021-11-12 | 索诺瓦有限公司 | Signal processing apparatus, system and method for processing audio signals |
CN113812173A (en) * | 2019-05-09 | 2021-12-17 | 索诺瓦有限公司 | Hearing device system and method for processing audio signals |
US11335332B2 (en) | 2019-09-03 | 2022-05-17 | Stmicroelectronics S.R.L. | Trigger to keyword spotting system (KWS) |
CN110675891A (en) * | 2019-09-25 | 2020-01-10 | 电子科技大学 | Voice separation method and module based on multilayer attention mechanism |
US11495215B1 (en) * | 2019-12-11 | 2022-11-08 | Amazon Technologies, Inc. | Deep multi-channel acoustic modeling using frequency aligned network |
WO2022076404A1 (en) * | 2020-10-05 | 2022-04-14 | The Trustees Of Columbia University In The City Of New York | Systems and methods for brain-informed speech separation |
US11875813B2 (en) | 2020-10-05 | 2024-01-16 | The Trustees Of Columbia University In The City Of New York | Systems and methods for brain-informed speech separation |
CN112528896A (en) * | 2020-12-17 | 2021-03-19 | 长沙理工大学 | SAR image-oriented automatic airplane target detection method and system |
CN112528896B (en) * | 2020-12-17 | 2024-05-31 | 长沙理工大学 | SAR image-oriented automatic aircraft target detection method and system |
CN113113000A (en) * | 2021-04-06 | 2021-07-13 | 重庆邮电大学 | Lightweight speech recognition method based on adaptive mask and grouping linear transformation |
CN113113000B (en) * | 2021-04-06 | 2022-05-13 | 重庆邮电大学 | Lightweight speech recognition method based on adaptive mask and grouping linear transformation |
CN113314136A (en) * | 2021-05-27 | 2021-08-27 | 西安电子科技大学 | Voice optimization method based on directional noise reduction and dry sound extraction technology |
CN115153563A (en) * | 2022-05-16 | 2022-10-11 | 天津大学 | Mandarin auditory attention decoding method and device based on EEG |
Also Published As
Publication number | Publication date |
---|---|
EP3469584B1 (en) | 2023-04-19 |
EP3469584A1 (en) | 2019-04-17 |
EP3469584A4 (en) | 2020-01-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11961533B2 (en) | Systems and methods for speech separation and neural decoding of attentional selection in multi-speaker environments | |
EP3469584B1 (en) | Neural decoding of attentional selection in multi-speaker environments | |
O’Sullivan et al. | Neural decoding of attentional selection in multi-speaker environments without access to clean sources | |
Han et al. | Speaker-independent auditory attention decoding without access to clean speech sources | |
CN110610719B (en) | Sound processing apparatus | |
Etard et al. | Neural speech tracking in the theta and in the delta frequency band differentially encode clarity and comprehension of speech in noise | |
Christiansen et al. | Prediction of speech intelligibility based on an auditory preprocessing model | |
Yang et al. | Speech reconstruction from human auditory cortex with deep neural networks. | |
Koning et al. | Speech onset enhancement improves intelligibility in adverse listening conditions for cochlear implant users | |
Duangudom et al. | Using auditory saliency to understand complex auditory scenes | |
Hazrati et al. | Reverberation suppression in cochlear implants using a blind channel-selection strategy | |
Soleymani et al. | SEDA: A tunable Q-factor wavelet-based noise reduction algorithm for multi-talker babble | |
Hosseini et al. | End-to-end brain-driven speech enhancement in multi-talker conditions | |
WO2021035067A1 (en) | Measuring language proficiency from electroencephelography data | |
Cantisani et al. | Neuro-steered music source separation with EEG-based auditory attention decoding and contrastive-NMF | |
KR102062454B1 (en) | Music genre classification apparatus and method | |
Gajecki et al. | A fused deep denoising sound coding strategy for bilateral cochlear implants | |
CN113178195B (en) | Speaker identification method based on sound-induced electroencephalogram signals | |
Zhang et al. | Sparsity-Driven EEG Channel Selection for Brain-Assisted Speech Enhancement | |
van der Heijden et al. | Joint population coding and temporal coherence link an attended talker's voice and location features in naturalistic multi-talker scenes. | |
Rao et al. | Automatic classification of healthy subjects and patients with essential vocal tremor using probabilistic source-filter model based noise robust pitch estimation | |
Hosseini et al. | Speaker-independent speech enhancement with brain signals | |
Li | Speech perception in a sparse domain | |
Xu et al. | An End-to-End EEG Channel Selection Method with Residual Gumbel Softmax for Brain-Assisted Speech Enhancement | |
Raghavan et al. | Improving auditory attention decoding by classifying intracranial responses to glimpsed and masked acoustic events |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 17813908 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2017813908 Country of ref document: EP Effective date: 20190114 |