US20170365271A1

US20170365271A1 - Automatic speech recognition de-reverberation

Info

Publication number: US20170365271A1
Application number: US15/388,323
Authority: US
Inventors: Adam Kupryjanow; Przemyslaw Maziewski; Lukasz Kurylo; Piotr Lasota
Original assignee: Intel Corp
Current assignee: Intel Corp
Priority date: 2016-06-15
Filing date: 2016-12-22
Publication date: 2017-12-21
Also published as: US10657983B2; CN109074816B; CN109074816A; US20170366897A1; EP3472834A1; US20170365255A1; WO2017218128A1; US20170365274A1; WO2017218129A1; EP3472834A4

Abstract

System and techniques for automatic speech recognition de-reverberation are described herein. A portion of an audio stream may be obtained. here, the portion of the audio stream is a proper subset of the audio stream. A filter may be created by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream. The filter may be applied to the audio stream to remove reverberation. The filtered version of the audio stream may then be provided to an audio stream consumer.

Description

CLAIM OF PRIORITY

This patent application claims the benefit of priority, under 35 U.S.C. §119, to U.S. Provisional Application Ser. No. 62/350,507, titled “FAR FIELD AUTOMATIC SPEECH RECOGNITION” and filed on Jun. 15, 2016, the entirety of which is hereby incorporated by reference herein.

TECHNICAL FIELD

Embodiments described herein generally relate to automatic speech recognition (ASR) and more specifically to ASR de-reverberation.

BACKGROUND

ASR involves a machine-based collection of techniques to understand human languages. ASR is interdisciplinary, often involving microphone, analog to digital conversion, frequency processing, database, and artificial intelligence technologies to convert the spoken word into textual or machine readable representations of not only what said (e.g., a transcript) but also what was meant (e.g., semantic understanding) by a human speaker. Far field ASR involves techniques to decrease a word error rate (WER) in utterances made a greater distance to a microphone, or microphone array, than traditionally accounted for in ASR processing pipelines. Such distance often decreases the signal to noise ratio (SNR) and thus increases WER in traditional ASR systems. As used herein, far field ASR involves distances more than half meter from the microphone.

BRIEF DESCRIPTION OF THE DRAWINGS

In the drawings, which are not necessarily drawn to scale, like numerals may describe similar components in different views. Like numerals having different letter suffixes may represent different instances of similar components. The drawings illustrate generally, by way of example, but not by way of limitation, various embodiments discussed in the present document.

FIG. 1 is an example of a smart home gateway housing, according to an embodiment.

FIG. 2 is a block diagram of an example high accuracy (HA) real-time de-reverberation device, according to an embodiment.

FIG. 3 is a block diagram of an example low latency (II) real-time de-reverberation device, according to an embodiment.

FIG. 4 is a block diagram of an example low complexity and high accuracy (LH) real-time de-reverberation device, according to an embodiment.

FIG. 5 is an example of a method for automatic speech recognition de-reverberation, according to an embodiment.

FIG. 6 is a block diagram illustrating an example of a machine upon which one or more embodiments may be implemented.

DETAILED DESCRIPTION

Embodiments and examples herein general described a number of systems, devices, and techniques for ASR de-reverberation. It is understood, however, that the systems, devices, and techniques are examples illustrating the underlying concepts.
Automatic speech recognition (ASR) often performs poorly if the user (or other sound source) is changing positions in the so-called far field (e.g., between one and five meters away from the microphone array). For example, for distances changing from 0.5 to 5 m the ASR word error rate (WER) for traditional techniques often grows from 4% to 18.5%. Due to issues of environmental acoustic conditions (e.g., reverberation or noise) that are exacerbated in far field ASR, near field (e.g., distances less than one half meter) ASR processing techniques perform poorly in the far field. Often far field conditions cause signals captured by microphones to be reverberant. In fact, when dealing with the far field ASR, reverberations may become the main parasitic factor decreasing ASR performance.
Reverb removal is not a trivial task because its characteristics depend on many factors, such as reverberation time (RT), distance between the sound source (e.g., user) and the microphone array, microphone array/device position, among other factors. Accordingly, even if a reverberation characteristic is known at one point in a room, for example, the characteristics may not be helpful to remove reverberation when a signal is captured in a different location (even if the location changed only a few millimeters).
In a reverberant scenario, many near-field pre-processing techniques that are used for acceptable ASR performance start to fail. What is needed is a de-reverberation technique to facilitate far field ASR performance improvements, such as, helping beam-formers to work properly in the reverberant conditions.
A framework for reducing reverberation is the generalized weighted prediction error (GWPE) de-reverberation technique. Although GWPE may lead to large WER improvements, it is expensive (e.g., in time, power, or device complexity) to compute. In tests, GWPE's real-time factor (RTF) is greater than 2.8 for a modern Intel® i7 processor on an eight microphone array. Here, the RTF is a ratio between the required time of computation and computed signal duration. Thus, when the RTF is higher than 1.0, the measured processing will not be real-time and the computation will take longer than the signal lasts. Because GWPE does not have an RTF less than or equal to one on most consumer devices, GWPE generally cannot be used in real-time solutions.
To address the issues with GWPE in real-time processing, several examples of systems and techniques are presented herein. Examples are able to run in real-time with, for example, an eight microphone array (such as that illustrated in FIG. 1). Some of these techniques improve WER at the five meter distance from 17.2% to 9.3% in clean conditions (e.g., without noises) and from 51.5% to 33.2% in noisy conditions. Removing reverberation helps improve ASR engine performance, beam-former techniques, and estimation of signal's source position, thus generally increasing the performance of ASR systems.
The devices or systems described below allow processing of an eight or less channel signal in real-time (e.g., RTF<=1.0). Further, in an example, some optimizations in the GWPE technique result in reduced computational complexity, allowing wider use of GWPE in a variety of devices with different computational capabilities.
FIG. 1 is an example of a device 100 including a smart home gateway housing 105, according to an embodiment. As illustrated, the circles atop the housing are lumens 110 behind which are housed microphones (as illustrated there are eight microphones). The dashed lines illustrate microphones in a linear arrangement 115 as well as in a circular arrangement 120. The device 100 includes a sampler 125, a signal processor 130, a multiplexer 140, and an interlink 145. In an example, the device includes a data store 135 to provide one or more buffers. All of these components are implemented in electronic hardware, such as that described below e.g., circuits).
The sampler 125 is arranged to obtain a portion of an audio stream. Here, the portion of the audio stream being a proper subset of the audio stream. To obtain the portion, the sampler 125 may receive the portion or may retrieve the portion from, for example, a microphone. In an example, obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period. In an example, the fixed time period (i.e. buffer length) is a second. In an example, the fixed time period is an audio frame. In an example, the audio frame length is thirty two milliseconds. The sampler 125 thus operates to obtain discrete audio samples.
The signal processor 130 is arranged to create a filter by applying GWPE to the portion of the audio stream. In an example, the signal processor 130 is arranged to calculate the frequency of domain data by applying an overlap and add (OLA) procedure combined with a Fast Fourier Transform (FFT) prior to creating the filter. A description of GWPE and its application in a number of examples is provided below with respect to FIGS. 2-4. Thus, instead of operating an entire utterance, GWPE is applied only to the portion obtained by the sampler 125 (e.g., a buffered signal). Reducing the size of the buffered segment of the signal reduces the computational complexity of traditional GWPE while still maintaining enough de-reverberation performance to benefit ASR processing. In an example, creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline. In an example, the first and second pipelines are arranged to execute in parallel on the device 100. Parallel execution here means actual simultaneous (e.g., not sequential as in a single core processor) execution of the pipelines. In an example, the signal processor 130 is also arranged to repeatedly create and apply the filter with a subsequent fixed time period (e.g., buttered portions of signal). Thus, the signal processor 130 creates and applies filters for respective samples obtained by the sampler 125.
In an example, creating the filter includes the signal processor 130 to combine a current GWPE application to the audio stream with a previously created filter. Here, GWPE application refers to the result of applying GWPE to the audio stream, in an example, combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term. In an example, combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding. In an example, the second scaling factor is between zero and one. In an example, the first scaling factor is one minus the second scaling factor.
The data store 135 provides a buffer to introduce a delay to the audio stream prior to applying the filter. This delay permits the processing buffer to fill with the required amount of signal. In an example, the delay is higher than 40 milliseconds and depends inter alia on the processor compute power.
The multiplexer 140 is arranged to apply the filter to the audio stream to remove reverberation. Thus, the multiplexer 140 accepts both the filter and the complex audio spectrum as inputs. The multiplexer 140 then applies the filter to the audio stream, calculates Inverse FFT(IFFT) and perform the OLA procedure to produce a signal with some (or all) of the reverberations removed. This signal may be referred to as a clean audio signal, filtered audio signal, or the like.
The interlink 145 is arranged to provide a filtered version of the audio stream to an audio stream consumer. Thus, the interlink 145 interfaces the present pipeline to other processing as part of a far field, or other, ASR system. Although far field ASR may greatly benefit from the device 100, the device 100 may also be helpful in other ASR situations, such as near field, to reduce reverberations.
FIGS. 2-4 illustrate several examples of configurations for a variety of the operations described above. These examples describe changes to a traditional GWPE application to meet the real-time processing used by many current devices while still maintaining a high degree of reverberation removal for ASR processing applications.
GWPE was originally designed in such a way that it required a whole recording of an utterance to properly design the de-reverberation filter. Thus, GWPE was not appropriate for real-time signal processing because processing must begin prior to an utterance's completion, denying the knowledge contained in the end of the recording that was used in the traditional GWPE implementation. To address the limitations of GWPE, three examples are described below. The first example is here labeled high accuracy (HA), the second example is here labeled low latency (LL), and the example is here labeled low-complexity with high accuracy (LH). The first and third examples (HA and LH) estimate the de-reverberation filter in a parallel to the channel processing. However, the role of the main thread is different between HA and LH. LL performs all calculations in a single (e.g., main) thread but uses estimates of the signal statistics instead of using the signal's actual (e.g., real) values.
To illustrate how these three examples improve on GWPE, the GWPE technique is introduced and then modifications to GWPE in each of the examples are explained.
GWPE operates in the frequency domain. GWPE operates on a signal with sample rate of 16 kHz and uses 32 ms frames with an 8 ms shift. GWPE treats every frequency bin independently. Thus, GWPE processes 257 independent frequency bins. As used herein, a frequency bin is represented by/and a number of the frame (e.g., index of the frame in a sample) by t. GWPE takes as an input of M channels and provides M channels of output. GWPE is a blind de-reverberation technique because all of the statistics needed by GWPE are obtained directly from the input signal. The GWPE de-reverberation operation is defined by the following equation:
$\begin{matrix} {\hat{X}}_{l} (t) = Y_{l} (t) - \sum_{τ = Δ}^{Δ + K_{l} - 1} {\hat{G}}_{l}^{*} (τ) Y_{l} (t - τ), \forall_{t \in T} & (1) \end{matrix}$
where:

- {circumflex over (X)}_l(t) is an estimate of a dry signal (e.g., without reverberations) for the time frame t from timespan T and frequency bin l,
- Y_l(t) is the signal captured by a microphone array;
- Ĝ*_l(τ) is an estimate of the de-reverberation filter;
- K_lis a filter length;
- Δ is a time delay; and
- {circumflex over (X)}_l(t), Y_l(t), and Ĝ*_l(τ) are vectors of length M.

Ĝ*_l(τ) may be estimated using the following:

- 1, initialization: Ĝ*_l(τ)=0 for all τ with Δ≦τ≦(Δ+K_l−1).
- 2. de-reverberation using equation (1).
- 3. spatial correlation matrix estimation:

(t)=E({circumflex over (X)} _l(t){circumflex over (X)}* _l(t))∀_tετ (2)

- 4. weighted correlation matrix/vector calculation:

$\begin{matrix} {\hat{R}}_{l} = \sum_{t \in τ}^{} ψ_{l} (\overline{t} - Δ) {(t)}^{- 1} ψ_{l}^{*} (\overline{t} - Δ) & (3) \\ \hat{r_{l}} = \sum_{t \in τ} ψ_{l} (\overline{t} - Δ) {(t)}^{- 1} Y_{l} (t) & (4) \end{matrix}$

- 5. filter update:

ĝ _l ={circumflex over (R)} _l ⁻¹ r _l (5)

- 6. convergence check

The complexity of the above technique is high. Given an eight channel input signal, the filter length is four. With 257 frequency bins, every filter update involves solving 257 linear equation systems, each of size 32 by 32 (R) (matrix) and 32 by 8 (r) (vector). From the perspective of signal quality it is important to update filter frequently enough to keep up (e.g., align) with reverberation characteristic changes.
GWPE updates the filter for an entire utterance. Thus, GWPE is inappropriate for working real-time working solutions due to:

- the latency it will introduce
- misalignment between the input signal and the filter in a case when the user moves during the utterance speaking
- resource requirements related to the memory needed to buffer the signal

FIG. 2 is a block diagram of an HA real-time de-reverberation device 205, according to an embodiment. To address some of the GWPE issues noted above, HA de-reverberation uses a short time span for filter estimation. For example, the time for estimation (T_est) may be equal to one second (1000 ms) Here, the filter is updated every T_est ms.
The elements of the HA device 205 include a main thread 215 and a parallel thread 210. The main thread 215 includes a storage device 235, a signal delay buffer 240, and a switcher 245. The parallel thread 210 includes a signal statistics calculator 220, a filter calculator 225, and a reverberation removal block 230. These components are all implemented in electronic hardware (e.g., circuits). The operations of the components proceed as follows:

- 1. The main thread 215 buffers the input signal in the storage device 235 and delays 240 samples at the beginning of processing.
- 2. When the buffer is filled (e.g., when 1000 ms of data has been collected), the GWPE procedure is performed in the parallel thread 210 using the technique described above, including calculating the signal statistics 220 and calculating the filter 225.
- 3. When the GWPE procedure is finished, the reverberation removal block 230 provides a signal with the reverberation removed to the switcher 245 to output. The filter update (e.g., blocks 220 and 225) may then be started for a new data chunk.

Thus, the processing proceeds with the latest (e.g., last) GWPE de-reverberation without imposing a delay for the entire utterance. An advantage of this technique includes its high accuracy of reverberation reduction. In an example, the delay introduced is:
ΔT=(1+RTF)T _est (6)
Where T_est is the time span used for the filter estimation and the RTF obtained from the GWPE implementation on the sample. In experimental results, the RTF approached 0.3 for this example.
FIG. 3 is a block diagram of an example LL real-time de-reverberation device 305, according to an embodiment. This additional example is here called the low-latency real-time de-reverberation. It uses an estimate of a weighted correlation matrix and vector ({circumflex over (R)}_land {circumflex over (τ)}_l) and does not wait for filter convergence. The device 305 includes a single thread 310 rather than the parallel threads described with respect to the I-IA real-time de-reverberation device 205 described above. The single thread includes a signal statistics estimation block 220 rather than the signal statistics calculation block described above, but otherwise also includes a filter calculator 320 and reverberation removal block 325. All of illustrated processing elements are implemented in electronic hardware. Operations of the device 305 proceed as follows:

- 1. initialization: Ĝ_l(t)=0 for all τ values with Δ≦τ≦Δ+K_l−1
- 2. De-reverberation using equation (1) (e.g., block 325)
- 3. Spatial correlation matrix estimation using equation (2) (e.g., block 315)
- 4. Weighted correlation matrix/vector calculation (e.g., block 320):

$\begin{matrix} {\hat{R}}_{l} (t) = {\begin{matrix} \sum_{t \in τ} ψ_{l (\overline{t} - Δ) {(t)}^{- 1} ψ_{l}^{*} (\overline{t} - Δ)}, for t = T_{\min} \\ α \cdot \hat{r_{l}} (t - a) + (a - α) \cdot \sum_{t \in τ} ψ_{l} (\overline{t} - Δ) {(t)}^{- 1} ψ_{l}^{*} (\overline{t} - Δ), for t > T_{\min} \end{matrix} & (7) \\ {\hat{r}}_{l} = {\begin{matrix} \sum_{t \in τ}^{} ψ_{l} (\overline{t} - Δ) {(t)}^{- 1} Y_{l} (t) \\ α \cdot \hat{r_{l}} (t - 1) + (1 - α) \cdot \sum_{t \in τ} ψ_{l} (\overline{t} - Δ) {(t)}^{- 1} Y_{l} (t) \end{matrix} & (8) \end{matrix}$

- 5. Filter update (e.g., block 320):

ĝ _l ={right arrow over (R)} _l ⁻¹ r _l , f or t≧T _min
where α is a smoothing factor (typically in a range from 0.9 to 0.999) and T_min is an initialization time span (typically in a range from 300 to 500 milliseconds ms)). This technique may introduce delay caused by the fast Fourier transform (FFT) and the overlap-add (OLA) procedures. For example, the delay is 40 ms when the frame size equals 32 ms and a frame shift of 8 ms. An additional benefit of this example includes lower memory requirements because calculating a new values of {right arrow over (R)}_land {circumflex over (r)}_lrequires only the last T signal frame to be buffered whereas the HA example may use T_est+T frames to be buffered. Additionally, because the filter is updated for every new frame, the precision of the de-reverberation filter may perform better than the HA example. Experimentally, a WER for clean speech of 10.5% for the HA example was achieved and a WER of 9.3% for the LL example was achieved.
FIG. 4 is a block diagram of an example LH real-time de-reverberation device 405, according to an embodiment. This example uses a similar foundation illustrated above between the threads of the HA example (FIG. 2), but uses the LL de-reverberation procedure albeit in a parallel thread. The delay introduced by this example may be set arbitrarily depending on the available computing power. Thus, the device 405 includes a main thread 430 and a parallel filter estimation thread 410. The main thread 430 includes a signal estimation block 435 which buffers results in a storage device 425. The main thread includes another storage area 445 (which may be the same physical device as storage device 425 in an example) to buffer the audio signal for use by the reverberation removal block 420 of the parallel thread 410. Additionally, the main thread 430 may include a delay 440 block to delay the audio signal into the switcher 450 to allow the filter to be processed prior to outputting the filtered signal. The parallel thread 410 includes the filter calculator 415 similar to that of the LL example to operates on the signal statistic estimates to produce the de-reverberation filter. This then is provided to the reverberation removal block 420 to perform the filtering.
The higher ΔT (e.g., the greater the time-frame sample) is set the lower the compute complexity becomes because the delay will often define how frequently the filter is updated. Thus, for ΔT=0, this example will be equivalent in compute complexity to the LL example described with respect to FIG. 3.
FIG. 5 is an example of a method 500 for automatic speech recognition de-reverberation, according to an embodiment. The operations of the method 500 are implemented and executed upon electronic hardware, such as that described above and below (e.g., circuits).
At operation 505, a portion of an audio stream is obtained. In an example, the portion of the audio stream is a proper subset of the audio stream. In an example, obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period. In an example, the fixed time period is a second. In an example, the fixed time period is an audio frame. In an example, the audio frame length is thirty two milliseconds. In an example, the method 500 may be extended to include repeating (e.g., repeatedly) creating the filter with a subsequent fixed time period.
At operation 510, a filter is created by applying GWPE to the portion of the audio stream. In an example, creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline. In an example, the first and second pipelines execute in parallel on a device.
In an example, creating the filter includes combining a current GWPE application to the audio stream with a previously created filter. In an example, combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term. In an example, combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding. In an example, the second scaling factor is between zero and one. In an example, the first scaling factor is one minus the second scaling factor.
At operation 515, the filter is applied to the audio stream to remove reverberation from the audio stream to produce a filtered version of the audio stream.
At operation 520, the filtered version of the audio stream is provided to an audio stream consumer.
The operations of the method 500 may be optionally extended to include introducing a delay to the audio stream prior to applying the filter. In an example, the delay is 40 milliseconds.
FIG. 6 illustrates a block diagram of an example machine 600 upon which any one or more of the techniques (e.g., methodologies) discussed herein may perform. In alternative embodiments, the machine 600 may operate as a standalone device or may be connected (e.g., networked) to other machines. In a networked deployment, the machine 600 may operate in the capacity of a server machine, a client machine, or both in server-client network environments. In an example, the machine 600 may act as a peer machine in peer-to-peer (P2P) (or other distributed) network environment. The machine 600 may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a mobile telephone, a web appliance, a network router, switch or bridge, or any machine capable of executing instructions (sequential or otherwise) that specify actions to be taken by that machine. Further, while only a single machine is illustrated, the term “machine” shall also be taken to include any collection of machines that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methodologies discussed herein, such as cloud computing, software as a service (SaaS), other computer cluster configurations.
Examples, as described herein, may include, or may operate by, logic or a number of components, or mechanisms. Circuitry is a collection of circuits implemented in tangible entities that include hardware (e.g., simple circuits, gates, logic, etc.). Circuitry membership may be flexible over time and underlying hardware variability. Circuitries include members that may, alone or in combination, perform specified operations when operating. In an example, hardware of the circuitry may be immutably designed to carry out a specific operation (e.g., hardwired). In an example, the hardware of the circuitry may include variably connected physical components (e.g., execution units, transistors, simple circuits, etc.) including a computer readable medium physically modified (e.g., magnetically, electrically, moveable placement of invariant massed particles, etc.) to encode instructions of the specific operation. In connecting the physical components, the underlying electrical properties of a hardware constituent are changed, for example, from an insulator to a conductor or vice versa. The instructions enable embedded hardware (e.g., the execution units or a loading mechanism) to create members of the circuitry in hardware via the variable connections to carry out portions of the specific operation when in operation. Accordingly, the computer readable medium is communicatively coupled to the other components of the circuitry when the device is operating. In an example, any of the physical components may be used in more than one member of more than one circuitry. For example, under operation, execution units may be used in a first circuit of a first circuitry at one point in time and reused by a second circuit in the first circuitry, or by a third circuit in a second circuitry at a different time.
Machine (e.g., computer system) 600 may include a hardware processor 602 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a hardware processor core, or any combination thereof), a main memory 604 and a static memory 606, some or all of which may communicate with each other via an interlink (e.g., bus) 608. The machine 600 may further include a display unit 610, an alphanumeric input device 612 (e.g., a keyboard), and a user interface (UI) navigation device 614 (e.g., a mouse). In an example, the display unit 610, input device 612 and UI navigation device 614 may be a touch screen display. The machine 600 may additionally include a storage device (e.g., drive unit) 616, a signal generation device 618 (e.g., a speaker), a network interface device 620, and one or more sensors 621, such as a global positioning system (UPS) sensor, compass, accelerometer, or other sensor. The machine 600 may include an output controller 628, such as a serial (e.g., universal serial bus (USB), parallel, or other wired or wireless (e.g., infrared (IR), near field communication (NFC), etc) connection to communicate or control one or more peripheral devices (e.g., a printer, card reader, etc.).
The storage device 616 may include a machine readable medium 622 on which is stored one or more sets of data structures or instructions 624 (e.g., software) embodying or utilized by any one or more of the techniques or functions described herein. The instructions 624 may also reside, completely or at least partially, within the main memory 604, within static memory 606, or within the hardware processor 602 during execution thereof by the machine 600. In an example, one or any combination of the hardware processor 602, the main memory 604, the static memory 606, or the storage device 616 may constitute machine readable media.
While the machine readable medium 622 is illustrated as a single medium, the term “machine readable medium” may include a single medium or multiple media (e.g., a centralized or distributed database, and/or associated caches and servers) configured to store the one or more instructions 624.
The term “machine readable medium” may include any medium that is capable of storing, encoding, or carrying instructions for execution by the machine 600 and that cause the machine 600 to perform any one or more of the techniques of the present disclosure, or that is capable of storing, encoding or carrying data structures used by or associated with such instructions. Non-limiting machine readable medium examples may include solid-state memories, and optical and magnetic media. In an example, a massed machine readable medium comprises a machine readable medium with a plurality of particles having invariant (e.g., rest) mass. Accordingly, massed machine-readable media are not transitory propagating signals. Specific examples of massed machine readable media may include: non-volatile memory, such as semiconductor memory devices (e.g., Electrically Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM)) and flash memory devices; magnetic disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.
The instructions 624 may further be transmitted or received over a communications network 626 using a transmission medium via the network interface device 620 utilizing any one of a number of transfer protocols (e.g., frame relay, internet protocol (IP), transmission control protocol (TCP), user datagram protocol (UDP), hypertext transfer protocol (HTTP), etc.). Example communication networks may include a local area network (LAN), a wide area network (WAN), a packet data network (e.g., the Internet), mobile telephone networks (e.g., cellular networks), Plain Old Telephone (POTS) networks, and wireless data networks (e.g., Institute of Electrical and Electronics Engineers (IEEE) 802.11 family of standards known as Wi-Fi®, IEEE 802.16 family of standards known as WiMax®), IEEE 802.15.4 family of standards, peer-to-peer (P2P) networks, among others. In an example, the network interface device 620 may include one or more physical jacks (e.g., Ethernet, coaxial, or phone jacks) or one or more antennas to connect to the communications network 626. In an example, the network interface device 620 may include a plurality of antennas to wirelessly communicate using at least one of single-input multiple-output (SIMO), multiple-input multiple-output (MIMO), or multiple-input single-output (MISO) techniques. The term “transmission medium” shall be taken to include any intangible medium that is capable of storing, encoding or carrying instructions for execution by the machine 600, and includes digital or analog communications signals or other intangible medium to facilitate communication of such software.

Additional Notes & Examples

Example 1 is a system for automatic speech recognition de-reverberation, the system comprising: a sampler to obtain a portion of an audio stream, the portion of the audio stream being a proper subset of the audio stream; a signal processor to create a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream; a multiplexer to apply the filter to the audio stream to remove reverberation; and an interlink to provide a filtered version of the audio stream to an audio stream consumer.
In Example 2, the subject matter of Example 1 optionally includes wherein the processor is in a first pipeline to create the filter and the multiplexer is in a second pipeline to apply the filter, the first and second pipelines arranged to execute in parallel.
In Example 3, the subject matter of any one or more of Examples 1-2 optionally include wherein, to obtain the portion of the audio stream, the sampler buffers the audio stream for a fixed time period.
In Example 4, the subject matter of Example 3 optionally includes wherein the fixed time period is a second.
In Example 5, the subject matter of any one or more of Examples 3-4 optionally include wherein the signal processor includes a loop to repetitively create the filter with subsequent fixed time periods.
In Example 6, the subject matter of any one or more of Examples 3-5 optionally include wherein the fixed time period is an audio frame.
In Example 7, the subject matter of Example 6 optionally includes wherein the audio frame length is thirty two milliseconds.
In Example 8, the subject matter of any one or more of Examples 1-7 optionally include wherein, to create the filter, the signal processor combines a current GWPE application to the audio stream with a previously created filter.
In Example 9, the subject matter of Example 8 optionally includes wherein, to combine the current GWPE application to the audio stream with a previously created filter, the signal processor adds the current GWPE application as a first term to the previously created filter as a second term.
In Example 10, the subject matter of Example 9 optionally includes wherein, to combine the current GWPE application to the audio stream with a previously created filter, the signal processor applies a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.
In Example 11, the subject matter of Example 10 optionally includes wherein the second scaling factor is between zero and one and wherein the first scaling factor is one minus the second scaling factor.
In Example 12, the subject matter of any one or more of Examples 1-11 optionally include a buffer to introduce a delay to the audio stream prior to applying the filter.
In Example 13, the subject matter of Example 12 optionally includes wherein the delay is eight milliseconds.
Example 14 is at least one machine readable medium including instructions for automatic speech recognition de-reverberation, the instructions, when executed by a machine, cause the machine to perform operations comprising: obtaining a portion of an audio stream, the portion of the audio stream being a proper subset of the audio stream; creating a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream; applying the filter to the audio stream to remove reverberation; and providing a filtered version of the audio stream to an audio stream consumer.
In Example 15, the subject matter of Example 14 optionally includes wherein creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline, the first and second pipelines executing in parallel on a device.
In Example 16, the subject matter of any one or more of Examples 14-15 optionally include wherein obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period.
In Example 17, the subject matter of Example 16 optionally includes wherein the fixed time period is a second.
In Example 18, the subject matter of any one or more of Examples 16-17 optionally include wherein the operations include repeating creating the filter with a subsequent fixed time period.
In Example 19, the subject matter of any one or more of Examples 16-18 optionally include wherein the fixed time period is an audio frame.
In Example 20, the subject matter of Example 19 optionally includes wherein the audio frame length is thirty two milliseconds.
In Example 21, the subject matter of any one or more of Examples 14-20 optionally include wherein creating the filter includes combining a current GWPE application to the audio stream with a previously created filter.
In Example 22, the subject matter of Example 21 optionally includes wherein combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term.
In Example 23, the subject matter of Example 22 optionally includes wherein combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.
In Example 24, the subject matter of Example 23 optionally includes wherein the second scaling factor is between zero and one and wherein the first scaling factor is one minus the second scaling factor.
In Example 25, the subject matter of any one or more of Examples 14-24 optionally include introducing a delay to the audio stream prior to applying the filter.
In Example 26, the subject matter of Example 25 optionally includes wherein the delay is eight milliseconds.
Example 27 is a device for automatic speech recognition de-reverberation, the device comprising: means for obtaining a portion of an audio stream, the portion of the audio stream being a proper subset of the audio stream; means for creating a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream; means for applying the filter to the audio stream to remove reverberation; and means for providing a filtered version of the audio stream to an audio stream consumer.
In Example 28, the subject matter of Example 27 optionally includes wherein the means for creating the filter occurs in a first pipeline and the means for applying the filter occurs in a second pipeline, the first and second pipelines executing in parallel on the device.
In Example 29, the subject matter of any one or more of Examples 27-28 optionally include wherein the means for obtaining the portion of the audio stream includes means for buffering the audio stream for a fixed time period.
In Example 30, the subject matter of Example 29 optionally includes wherein the fixed time period is a second.
In Example 31, the subject matter of any one or more of Examples 29-30 optionally include means for repeating creating the filter with a subsequent fixed time period.
In Example 32, the subject matter of any one or more of Examples 29-31 optionally include wherein the fixed time period is an audio frame.
In Example 33, the subject matter of Example 32 optionally includes wherein the audio frame length is thirty two milliseconds.
In Example 34, the subject matter of any one or more of Examples 27-33 optionally include wherein the means for creating the filter includes means for combining a current GWPE application to the audio stream with a previously created filter.
In Example 35, the subject matter of Example 34 optionally includes wherein the means for combining the current GWPE application to the audio stream with a previously created filter includes means for adding the current GWPE application as a first term to the previously created filter as a second term.
In Example 36, the subject matter of Example 35 optionally includes wherein the means for combining the current GWPE application to the audio stream with a previously created filter includes means for applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.
In Example 37, the subject matter of Example 36 optionally includes wherein the second scaling factor is between zero and one and wherein the first scaling factor is one minus the second scaling factor.
In Example 38, the subject matter of any one or more of Examples 27-37 optionally include wherein the means for applying the filter to the audio stream includes means for introducing a delay to the audio stream prior to applying the filter.
In Example 39, the subject matter of Example 38 optionally includes wherein the delay is eight milliseconds.
Example 40 is a method for automatic speech recognition de-reverberation, the method comprising: obtaining a portion of an audio stream; the portion of the audio stream being a proper subset of the audio stream; creating a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream; applying the filter to the audio stream to remove reverberation; and providing a filtered version of the audio stream to an audio stream consumer.
In Example 41, the subject matter of Example 40 optionally includes wherein creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline; the first and second pipelines executing in parallel on a device.
In Example 42, the subject matter of any one or more of Examples 40-41 optionally include wherein obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period.
In Example 43, the subject matter of Example 42 optionally includes wherein the fixed time period is a second.
In Example 44, the subject matter of any one or more of Examples 42-43 optionally include repeating creating the filter with a subsequent fixed time period.
In Example 45, the subject matter of any one or more of Examples 42-44 optionally include wherein the fixed time period is an audio frame.
In Example 46, the subject matter of Example 45 optionally includes wherein the audio frame length is thirty two milliseconds.
In Example 47, the subject matter of any one or more of Examples 40-46 optionally include wherein creating the filter includes combining a current GWPE application to the audio stream with a previously created filter.
In Example 48, the subject matter of Example 47 optionally includes wherein combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term.
In Example 49, the subject matter of Example 48 optionally includes wherein combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.
In Example 50, the subject matter of Example 49 optionally includes wherein the second scaling factor is between zero and one and wherein the first scaling factor is one minus the second scaling factor.
In Example 51, the subject matter of any one or more of Examples 40-50 optionally include introducing a delay to the audio stream prior to applying the filter.
In Example 52, the subject matter of Example 51 optionally includes wherein the delay is eight milliseconds.
Example 53 is a system comprising means to perform any of the methods 40-52.
Example 54 is at least one machine readable medium including instructions that, when executed by a machine, cause the machine to perform any of methods 40-52.
Example 55 is at least one machine readable medium including instructions for de-reverberation of an audio signal, the instructions, when executed by a machine, causing the machine to perform operations comprising: performing Generalized Weighted Prediction Error (GWPE) in a first pipeline; and performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 56, the subject matter of Example 55 optionally includes buffering the audio signal in a buffer; providing contents of the buffer every second to the first pipeline; and clearing the buffer after providing the contents.
In Example 57, the subject matter of any one or more of Examples 55-56 optionally include wherein the first pipeline includes iteratively: calculating signal statistics; calculating a de-reverb filter; and applying the de-reverb filter to remove reverberation.
Example 58 is a method for de-reverberation of an audio signal, the method comprising: performing Generalized Weighted Prediction Error (GWPE) in a first pipeline; and performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 59, the subject matter of Example 58 optionally includes buffering the audio signal in a buffer; providing contents of the buffer every second to the first pipeline; and clearing the buffer after providing the contents.
In Example 60, the subject matter of any one or more of Examples 58-59 optionally include wherein the first pipeline iteratively includes: calculating signal statistics; calculating a de-reverb filter; and applying the de-reverb filter to remove reverberation.
Example 61 is a system comprising means to perform any of the methods 58-60.
Example 62 is at least one machine readable medium including instructions that, when executed by a machine, cause the machine to perform any of methods 58-60.
Example 63 is a system for de-reverberation of an audio signal, the system comprising: means for performing Generalized Weighted Prediction Error (GWPE) in a first pipeline; and means for performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 64, the subject matter of Example 63 optionally includes means for buffering the audio signal in a buffer; means for providing contents of the butter every second to the first pipeline; and means for clearing the buffer after providing the contents.
In Example 65, the subject matter of any one or more of Examples 63-64 optionally include wherein the first pipeline includes means for iteratively: calculating signal statistics; calculating a de-reverb filter; and applying the de-reverb filter to remove reverberation.
Example 66 is at least one machine readable medium including instructions for de-reverberation of an audio signal, the instructions, when executed by a machine, causing the machine to perform operations comprising: estimating signal statistics for an audio signal; performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; estimating a spatial correlation matrix; creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and updating a de-reverb filter with the weighted matrix and vector.
In Example 67, the subject matter of Example 66 optionally includes are performed inline to other audio signal processing.
In Example 68, the subject matter of any one or more of Examples 66-67 optionally include wherein only one signal frame is buffered at a time to input into the operations.
Example 69 is a method for de-reverberation of an audio signal, the method comprising: estimating signal statistics for an audio signal; performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; estimating a spatial correlation matrix; creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and updating a de-reverb filter with the weighted matrix and vector.
In Example 70, the subject matter of Example 69 optionally includes are performed inline to other audio signal processing.
In Example 71, the subject matter of any one or more of Examples 69-70 optionally include wherein only one signal frame is buffered at a time to input into the operations.
Example 72 is a system comprising means to perform any of the methods 69-71.
Example 73 is at least one machine readable medium including instructions that, when executed by a machine, cause the machine to perform any of methods 69-71.
Example 74 is a system for de-reverberation of an audio signal, the system comprising: means for estimating signal statistics for an audio signal; means for performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; means for estimating a spatial correlation matrix; means for creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and means for updating a de-reverb filter with the weighted matrix and vector.
In Example 75, the subject matter of Example 74 optionally includes are performed inline to other audio signal processing.
In Example 76, the subject matter of any one or more of Examples 74-75 optionally include wherein only one signal frame is buffered at a time to input into the operations.
Example 77 is at least one machine readable medium including instructions for de-reverberation of an audio signal, the instructions, when executed by a machine, causing the machine to perform operations comprising: performing de-reverb filter updating in a first pipeline; and performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 78, the subject matter of Example 77 optionally includes wherein the de-reverb filter updating includes: estimating signal statistics for the audio signal; performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; estimating a spatial correlation matrix; creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and updating a de-reverb filter with the weighted matrix and vector.
In Example 79, the subject matter of any one or more of Examples 77-78 optionally include wherein only one signal frame is buffered at a time to input into the first pipeline.
Example 80 is a method for de-reverberation of an audio signal, the method comprising: performing de-reverb filter updating in a first pipeline; and performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 81, the subject matter of Example 80 optionally includes wherein the de-reverb filter updating includes: estimating signal statistics for the audio signal; performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; estimating a spatial correlation matrix; creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and updating a de-reverb filter with the weighted matrix and vector.
In Example 82, the subject matter of any one or more of Examples 80-81 optionally include wherein only one signal frame is buffered at a time to input into the first pipeline.
Example 83 is a system comprising means to perform any of the methods 80-82.
Example 84 is at least one machine readable medium including instructions that, when executed by a machine, cause the machine to perform any of methods 80-82.
Example 85 is a system for de-reverberation of an audio signal, the system comprising: means for performing dc-reverb filter updating in a first pipeline; and means for performing signal processing in a second pipeline, the second pipeline and first pipeline executing in parallel, the second pipeline applying the output of the first pipeline to remove reverberation in an audio signal processed by the second pipeline.
In Example 86, the subject matter of Example 85 optionally includes wherein the de-reverb filter updating includes: means for estimating signal statistics for the audio signal; means for performing Generalized Weighted Prediction Error (GWPE) using the estimated signal statistics; means for estimating a spatial correlation matrix; means for creating weighted matrix and vector inputs from the spatial correlation matrix estimation; and means for updating a de-reverb filter with the weighted matrix and vector.
In Example 87, the subject matter of any one or more of Examples 85-86 optionally include wherein only one signal frame is buffered at a time to input into the first pipeline.
The above detailed description includes references to the accompanying drawings, which form a part of the detailed description. The drawings show, by way of illustration, specific embodiments that may be practiced. These embodiments are also referred to herein as “examples.” Such examples may include elements in addition to those shown or described. However, the present inventors also contemplate examples in which only those elements shown or described are provided. Moreover, the present inventors also contemplate examples using any combination or permutation of those elements shown or described (or one or more aspects thereof), either with respect to a particular example (or one or more aspects thereof), or with respect to other examples (or one or more aspects thereof) shown or described herein.
All publications, patents, and patent documents referred to in this document are incorporated by reference herein in their entirety, as though individually incorporated by reference. In the event of inconsistent usages between this document and those documents so incorporated by reference, the usage in the incorporated reference(s) should be considered supplementary to that of this document; for irreconcilable inconsistencies, the usage in this document controls.
In this document, the terms “a” or “an” are used, as is common in patent documents, to include one or more than one, independent of any other instances or usages of “at least one” or “one or more.” In this document, the term “or” is used to refer to a nonexclusive or, such that “A or B” includes “A but not B,” “B but not A,” and “A and B,” unless otherwise indicated. In the appended claims, the terms “including” and “in which” are used as the plain-English equivalents of the respective terms “comprising” and “wherein.” Also, in the following claims, the terms “including” and “comprising” are open-ended, that is, a system, device, article, or process that includes elements in addition to those listed after such a term in a claim are still deemed to fall within the scope of that claim. Moreover, in the following claims, the terms “first,” “second,” and “third,” etc. are used merely as labels, and are not intended to impose numerical requirements on their objects.
The above description is intended to be illustrative, and not restrictive. For example, the above-described examples (or one or more aspects thereof) may be used in combination with each other. Other embodiments may be used, such as by one of ordinary skill in the art upon reviewing the above description. The Abstract is to allow the reader to quickly ascertain the nature of the technical disclosure and is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. Also, in the above Detailed. Description, various features may be grouped together to streamline the disclosure. This should not be interpreted as intending that an unclaimed disclosed feature is essential to any claim. Rather, inventive subject matter may lie in less than all features of a particular disclosed embodiment. Thus, the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separate embodiment. The scope of the embodiments should be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

What is claimed is:

1. A system for automatic speech recognition de-reverberation, the system comprising:

a sampler to obtain a portion of an audio stream, the portion of the audio stream being a proper subset of the audio stream;

a signal processor to create a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream;

a multiplexer to apply the filter to the audio stream to remove reverberation; and

an interlink to provide a filtered version of the audio stream to an audio stream consumer.

2. The system of claim 1, wherein the signal processor is in a first pipeline to create the filter and the multiplexer is in a second pipeline to apply the filter, the first and second pipelines arranged to execute in parallel.

3. The system of claim 1, wherein, to obtain the portion of the audio stream, the sampler buffers the audio stream for a fixed time period.

4. The system of claim 3, wherein the signal processor includes a loop to repetitively create the filter with subsequent fixed time periods.

5. The system of claim 3, wherein fixed time period is an audio frame.

6. The system of claim 1, wherein, to create the filter, the signal processor combines a current GWPE application to the audio stream with a previously created filter.

7. The system of claim 6, wherein, to combine the current GWPE application to the audio stream with a previously created filter, the signal processor adds the current GWPE application as a first term to the previously created filter as a second term.

8. The system of claim 7, wherein, to combine the current GWPE application to the audio stream with a previously created filter, the signal processor applies a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.

9. At least one machine readable medium including instructions for automatic speech recognition de-reverberation, the instructions, when executed by a machine, cause the machine to perform operations comprising:

obtaining a portion of an audio stream, the portion of the audio stream being a proper subset of the audio stream;

creating a filter by applying Generalized Weighted Prediction Error (GWPE) to the portion of the audio stream;

applying the filter to the audio stream to remove reverberation; and

providing a filtered version of the audio stream to an audio stream consumer.

10. The at least one machine readable medium of claim 9, wherein creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline, the first and second pipelines executing in parallel on a device.

11. The at least one machine readable medium of claim 9, wherein obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period.

12. The at least one machine readable medium of claim 11, wherein the operations include repeating creating the filter with a subsequent fixed time period.

13. The at least one machine readable medium of claim 11, wherein the fixed time period is an audio frame.

14. The at least one machine readable medium of claim 9, wherein creating the filter includes combining a current GWPE application to the audio stream with a previously created filter.

15. The at least one machine readable medium of claim 14, wherein combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term.

16. The at least one machine readable medium of claim 15, wherein combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.

17. A method for automatic speech recognition de-reverberation, the method comprising:

applying the filter to the audio stream to remove reverberation; and

providing a filtered version of the audio stream to an audio stream consumer.

18. The method of claim 17, wherein creating the filter occurs in a first pipeline and applying the filter occurs in a second pipeline, the first and second pipelines executing in parallel on a device.

19. The method of claim 17, wherein obtaining the portion of the audio stream includes buffering the audio stream for a fixed time period.

20. The method of claim 19, comprising repeating creating the filter with a subsequent fixed time period.

21. The method of claim 19, wherein the fixed time period is an audio frame.

22. The method of claim 17, wherein creating the filter includes combining a current GWPE application to the audio stream with a previously created filter.

23. The method of claim 22, wherein combining the current GWPE application to the audio stream with a previously created filter includes adding the current GWPE application as a first term to the previously created filter as a second term.

24. The method of claim 23, wherein combining the current GWPE application to the audio stream with a previously created filter includes applying a first scaling factor to the first term and a second scaling factor to the second term prior to the adding.