US20140142946A1 - System and method for voice transformation - Google Patents
System and method for voice transformation Download PDFInfo
- Publication number
- US20140142946A1 US20140142946A1 US13/625,317 US201213625317A US2014142946A1 US 20140142946 A1 US20140142946 A1 US 20140142946A1 US 201213625317 A US201213625317 A US 201213625317A US 2014142946 A1 US2014142946 A1 US 2014142946A1
- Authority
- US
- United States
- Prior art keywords
- voice
- timbre
- new
- vectors
- signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 38
- 230000009466 transformation Effects 0.000 title claims abstract description 18
- 239000013598 vector Substances 0.000 claims abstract description 77
- 238000001228 spectrum Methods 0.000 claims abstract description 47
- 230000006870 function Effects 0.000 claims abstract description 26
- 230000015572 biosynthetic process Effects 0.000 claims abstract description 18
- 238000003786 synthesis reaction Methods 0.000 claims abstract description 18
- 230000001131 transforming effect Effects 0.000 claims abstract 5
- 238000012545 processing Methods 0.000 claims description 16
- 238000001514 detection method Methods 0.000 claims description 4
- 230000003595 spectral effect Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 description 8
- 230000004048 modification Effects 0.000 description 6
- 238000012986 modification Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 210000004704 glottis Anatomy 0.000 description 5
- 230000011218 segmentation Effects 0.000 description 5
- 238000000926 separation method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000000737 periodic effect Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 241001672694 Citrus reticulata Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 230000002996 emotional effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 210000004072 lung Anatomy 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 239000000523 sample Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 210000001260 vocal cord Anatomy 0.000 description 1
- 230000001755 vocal effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/04—Segmentation; Word boundary detection
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L13/00—Speech synthesis; Text to speech systems
- G10L13/08—Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L13/00—Speech synthesis; Text to speech systems
- G10L13/02—Methods for producing synthetic speech; Speech synthesisers
- G10L13/04—Details of speech synthesis systems, e.g. synthesiser structure or memory management
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
Definitions
- the present invention generally relates to voice transformation, in particular to voice transformation using orthogonal functions, and its applications in speech synthesis and automatic speech recognition.
- Voice transformation involves parameterization of a speech signal into a mathematical format which can be extensively manipulated such that the properties of the original speech, for example, pitch, speed, relative length of phones, prosody, and speaker identity, can be changed, but still sound natural.
- a straightforward application of voice transformation is singing synthesis. If the new parametric representation is successfully demonstrated to work well in voice transformation, it can be used for speech synthesis and automatic speech recognition.
- Speech synthesis involves the use of a computer-based system to convert a written document into audible speech.
- a good TTS system should generate natural, or human-like, and highly intelligible speech.
- the rule-based TTS systems or the formant synthesizers, were used. These systems generate intelligible speech, but the speech sounds robotic, and unnatural.
- ASR Automatic speech recognition
- the present invention is directed to a novel mathematical representation of the human voice as a timbre vector, together with a method of parameterizing speech into a timbre vector, and a method to recover human voice from a series of timbre vectors with variations.
- a speech signal is first segmented into non-overlapping frames using the glottal closure moment information. Using Fourier analysis, the speech signal in each frame is converted into amplitude spectrum, then Laguerre functions (based on a set of orthogonal polynomials) are used to convert the amplitude spectrum into a unit vector characteristic to the instantaneous timbre.
- a timbre vector is formed along with voicedness index, frame duration, and an intensity parameter. Because of the accuracy of the system and method and the complete separation of prosody and timbre, a variety of voice transformation operations can be applied, and the output voice is natural. A straightforward application of voice transformation is singing synthesis.
- One difference of the current invention from all previous methods is that the frames, or processing units, are non-overlapping, and do not require a window function.
- All previous parameterization methods including linear prediction confidents, sinusoidal models, mel-cepstral coefficients and time-domain pitch synchronized overlap add methods rely on overlapping frames requiring a window function (such as Hamming window, Hann window, cosine window, triangular window, Gaussian window, etc.) and a shift time which is smaller than the duration of the frame, which makes an overlap.
- a window function such as Hamming window, Hann window, cosine window, triangular window, Gaussian window, etc.
- An important application of the inventive parametric representation is speech synthesis.
- the speech segments can be modified to the prosodic requirements and regenerate an output speech with high quality.
- the synthesized speech can have different speaker identity (baby, child, male, female, giant, etc), base pitch (up to three octaves), speed (up to 10 times), and various prosodic variations (calm, emotional, up to shouting).
- the timbre vector method disclosed in the present invention can be used to build high-quality speech synthesis systems using a compact speech database.
- Another important application of the inventive parametric representation of speech signal is to serve as the acoustic signal format to improve the accuracy of automatic speech recognition.
- the timbre vector method disclosed in the present invention can greatly improve the accuracy of automatic speech recognition.
- FIG. 1 is a block diagram of a voice transformation systems using timbre vectors according to an exemplary embodiment of the present invention.
- FIG. 2 is an explanation of the basic concept of parameterization according to an exemplary embodiment of the present invention.
- FIG. 3 is the process of segmenting the PCM data according to an exemplary embodiment of the present invention.
- FIG. 4 is a plot of the Laguerre functions according to an exemplary embodiment of the present invention.
- FIG. 5 is the data structure of a timbre vector according to an exemplary embodiment of the present invention.
- FIG. 6 is the binomial interpolation of timbre vectors according to an exemplary embodiment of the present invention.
- FIG. 7 is a block diagram of a speech synthesis system using timbre vectors according to an exemplary embodiment of the present invention.
- FIG. 8 is a block diagram of an automatic speech recognition system using timbre vectors according to an exemplary embodiment of the present invention.
- Various exemplary embodiments of the present invention are implemented on a computer system including one or more processors and one or more memory units.
- steps of the various methods described herein are performed on one or more computer processors according to instructions encoded on a computer-readable medium.
- FIG. 1 is a block diagram of the voice transformation system according to an exemplary embodiment of the present invention.
- the source is the voice from a speaker 101 .
- PCM Pulse Code Modulation
- the PCM signal 103 is then segmented by segmenter 104 into frames 105 , according to segment points 110 .
- the second one is to use a glottal closure instants detection unit 108 to generate GCI from the voice waveform.
- the glottal closure instants (GCI) 107 and the voice signal (PCM) 103 are sent to a processing unit 109 , to generate a complete set of segment points 110 .
- FIG. 3 The details of this process is shown in FIG. 3 .
- the voice signal in each frame 105 proceeds through a Fourier analysis unit 111 to generate amplitude spectrum 112 .
- the amplitude spectrum 112 proceeds through an orthogonal transform unit 113 to generate timbre vectors 114 .
- Laguerre functions are the most appropriate mathematical functions for converting the amplitude spectrum into a compact and convenient form (see FIG. 4 ).
- Data structure of a timbre vector is shown in FIG. 5 .
- a number of voice manipulations can be made according to specifications 115 by voice manipulator 116 , so as to generate new timbre vectors 117 , then the voice can be regenerated using the new timbre vectors 117 .
- Laguerre transform 118 is used to regenerate amplitude spectrum 119 ; the phase generator 120 (based on Kramers-Kronig relations) is used to generate phase spectrum 121 ; FFT (Fast Fourier Transform) 122 is used to generate an elementary acoustic wave 123 , from the amplitude spectrum and phase spectrum; then those elementary acoustic waves 123 are superposed according to the timing information 124 in the new timbre vectors, each one is delayed by the time of frame duration 125 of the previous frame.
- the output wave in electric form then drives a loudspeaker 126 to produce an output voice 127 .
- FIG. 2 shows the process of speech generation, particularly the generation of voiced sections, and the properties of the PCM and EGG signals.
- Air flow 201 comes from the lungs to the opening between the two vocal cords, or glottis, 202 . If the glottis is constantly open, there is a constant air flow 203 , but no voice signal is generated. At the instant the glottis closes, or a glottal closure occurs, which is always very rapid due to the Bernoulli effect, the inertia of the moving air in the vocal track 204 generates a d'Alembert wave front, then excites an acoustic resonance.
- the actions of the glottis is monitored by the signals from a electroglottograph (EGG) 205 .
- EGG electroglottograph
- the instrument When there is a glottal closure, the instrument generates a sharp peak in the derivative of the EGG signal, as shown as 207 in FIG. 2 .
- a microphone 206 is placed near the mouth to generate a signal, typically a Pulse Code Modulation signal, or PCM, as shown in 209 in FIG. 2 . If the glottis remains closed after a closure, as shown as shown as 208 , then the acoustic excitation sustains, as shown as 210 .
- FIG. 3 shows the details of processing unit 109 to generate the segmentation points.
- the input data is the PCM signal 301 - 303 and EGG signal 304 , produced by the source speaker 101 .
- EGG signal 304 When there are clear peaks in the EGG signal, such as 304 , corresponding to PCM signal 301 , those peaks are selected as the segmentation points 305 .
- the segmentation points are generated by comparing the waveform 302 with the neighboring ones 301 , and if the waveform 302 is still periodic, then segmentation points 306 are generated at the same intervals as the segmentation points 305 . If the signal is no longer periodic, such as 303 , the PCM is segmented according to points 307 into frames with an equal interval, here 5 msec. Therefore, the entire PCM signal is segmented into frames.
- the values of the voice signal at two adjacent closure moments may not match.
- the following is an algorithm that may be used to match the ends. Let the number of sampling points between two adjacent glottal closures be N, and the original voice signal be x 0 (n). The smoothed signal x(n) in a small interval 0 ⁇ n ⁇ M is defined as
- x ⁇ ( N - n ) x 0 ⁇ ( N - n ) ⁇ n M + x 0 ⁇ ( - n ) ⁇ M - n M .
- ⁇ n ⁇ ( x ) n ! ( n + k ) ! ⁇ ⁇ - x / 2 ⁇ x k / 2 ⁇ L n ( k ) ⁇ ( x ) ,
- the norm of the vector C is the intensity parameter I,
- ⁇ ⁇ ( ⁇ ) - 1 ⁇ ⁇ ⁇ lim ⁇ ⁇ 0 ⁇ ⁇ [ ⁇ - ⁇ ⁇ - ⁇ ⁇ ln ⁇ ⁇ A ⁇ ( ⁇ ′ ) ⁇ ′ - ⁇ ⁇ ⁇ ⁇ ⁇ ′ + ⁇ ⁇ + ⁇ ⁇ ⁇ ln ⁇ ⁇ A ⁇ ( ⁇ ′ ) ⁇ ′ - ⁇ ⁇ ⁇ ⁇ ⁇ ′ ]
- the output wave for a frame, the acoustic exciton can be calculated from the amplitude spectrum A( ⁇ ) and the phase spectrum ⁇ ( ⁇ ),
- x ⁇ ( t ) ⁇ 0 ⁇ ⁇ A ⁇ ( ⁇ ) ⁇ cos ⁇ ( ⁇ ⁇ ⁇ t - ⁇ ⁇ ( ⁇ ) ) ⁇ ⁇ ⁇ .
- FIG. 4 shows the Laguerre function. After proper scaling, twenty-nine Laguerre functions are used on the frequency scale 401 of 0 to 11 kHz.
- the first Laguerre function 402 actually probes the first formant.
- the resolution in the low-frequency range is successively improved; and extended to the high-frequency range 404 . Because of the accuracy scaling, it makes an accurate but concise representation of the spectrum.
- FIG. 5 shows the data structure of a timbre vector including the voicedness index (V) 501 , the frame duration (T) 502 , the intensity parameter (I) 503 , and the normalized Laguerre coefficients 504 .
- Timbre interpolation The unit vector of Laguerre coefficients varies slowly with frames. It can be interpolated for reduced number of frames or extended number of frames for any section of voice to produce natural sounding speech of arbitrary temporal variations. For example, the speech can be made very fast but still recognizable by a blind person.
- Timbre fusing By connecting two sets of timbre vectors of two different phonemes and smear-averaging over the juncture, a natural-sounding transition is generated. Phoneme assimilation may be automatically produced. By connecting a syllable ended with [g] with a syllable started with [n], after fusing, the sound [n] is automatically assimilated into [ng].
- FIG. 6 shows the principles of the timbre fusing operation.
- Original timbre vectors from the first phoneme 601 include timbre vectors A, B, and C.
- Original timbre vectors from the second phoneme 602 include timbre vectors D and E.
- the output timbre vectors 603 through 607 are weighed averages from the original timbre vectors. For example, output timbre vector D′ is generated from timbre vector C, D, and E using the binomial coefficients 1, 2, and 1; output timbre vector C′ is generated from original timbre vectors A, B, C, D, and E using the binomial coefficients 1, 4, 6, 4, and 1.
- the state-of-the-art technology for pitch modification of speech signal is the time-domain pitch-synchronized overlap-add (TD-PSOLA) method, which can change pitch from ⁇ 30% to +50%. Otherwise the output would sound unnatural.
- TD-PSOLA time-domain pitch-synchronized overlap-add
- pitch can be easily modified by changing the time of separation T, then using timbre interpolation to compensate speed.
- Natural sounding speech can be produced with pitch modifications as large as three octaves.
- intensity parameter I is a property of a frame, it can be changed to produce any stress pattern required by prosody input.
- the head size can be changed.
- the voice of an average adult speaker can be changed to that of a baby, a child, a woman, a man, or a giant.
- high-quality speech synthesizers with a compact database can be constructed using the parametric representation based on timbre vectors (see FIG. 7 ).
- the speech synthesis system has two major parts: database building part 101 (the left-hand side of FIG. 7 ), and the synthesis part 121 (right-hand side of FIG. 7 ).
- a source speaker 702 reads a prepared text.
- the voice is recorded by a microphone to become the PCM signal 703 .
- the glottal closure signal is recorded by an electroglottograph (EGG) to become EGG signal 704 .
- EGG electroglottograph
- the origin and properties of those signals are shown in FIG. 2 .
- the EGG signal and the PCM signal are used by the processing unit 705 to generate a set of segment points 706 .
- the details of the segmenting process, or the function of the processing unit, is shown in FIG. 3 .
- the PCM signal is segmented by the segmenter 707 into frames 708 using the segment points 706 . Each frame is processed by a unit of Fourier analysis 709 to generate amplitude spectrum 710 .
- the amplitude spectrum of each frame is then processed using a Laguerre transform unit 711 to become a unit vector, representing the instantaneous timbre of that frame, to become the basis of timbre vectors 712 .
- the Laguerre functions are shown in FIG. 4 .
- the structure of the timbre vector is shown in FIG. 5 .
- the timbre vectors of various units of speech such as, for example, phonemes, diphones, demisyllables, syllables, words and even phrases, are then stored in the speech database 720 .
- the input text 722 together with synthesis parameters 723 are fed into the frontend 724 .
- Detailed instructions about the phonemes, intensity and pitch values 725 , for generating the desired speech are generated, then input to a processing unit 726 .
- the processing unit 726 selects timbre vectors from the database 720 , then converts the selected timbre vectors to a new series of timbre vectors 727 according to the instructions from the process unit 726 , and using timbre fusing if necessary (see FIG. 6 ).
- Each timbre vector is converted into an amplitude spectrum 729 by Laguerre transform unit 728 .
- the phase spectrum 731 is generated from the amplitude spectrum 729 by phase generator 730 using a Kramers-Kronig relations algorithm.
- the amplitude spectrum 729 and the phase spectrum 731 are sent to a FFT (Fast Fourier Transform) unit 732 , to generate an elementary acoustic wave 733 .
- FFT Fast Fourier Transform
- Those elementary acoustic waves 733 are than superposed by the superposition unit 735 according to the timing information 734 provided by the new timbre vectors 727 , to generate the final result, output speech signal 736 .
- the parametric representation of human voice in terms of timbre vectors can also be used as the basis of automatic speech recognition systems.
- the most widely used acoustic features, or parametric representation of human speech in automatic speech recognition is the mel-cepstrum.
- the speech signal is segmented into frames of fixed length, typically 20 msec, with a window, typically Hann window or Hamming window, and a shift of 10 msec.
- Those parametric representations are crude and inaccurate. Features that cross the phoneme borders occur very often.
- the parametric representation based on timbre vectors is more accurate.
- a well-behaved timbre distance ⁇ between two frames can be defined as
- c (1) n and c (2) n are elements of the normalized Laguerre coefficients of the two timbre vectors (see FIG. 5 ).
- the distance is less than 0.1.
- the distance is 0.1 to 0.6.
- vowels and unvoiced consonants are well separated.
- the intensity parameter I silence is well separated from real sound.
- pitch is an important parameter (see, for example, U.S. Pat. No. 5,751,905 and U.S. Pat. No. 6,510,410).
- the frame duration T provides a very accurate measure of pitch (see FIG. 5 ). Therefore, using parametric representation based on timbre vectors, the accuracy of speech recognition can be greatly improved.
- FIG. 8 shows a block diagram of an automatic speech recognition system based on timbre vectors.
- the first half of the procedure, converting speech signal into timbre vectors, is similar to step 102 through step 114 of FIG. 1 for voice transformation.
- the voice from a speaker 801 is recorded in the computer as PCM signal 803 .
- the PCM signal 803 is then segmented by segmenter 804 into frames 805 , according to segment points 810 .
- the first one is to use an electroglottograph (EGG) 806 to detect the glottal closure instants (GCI) 807 directly (see FIG. 2 ).
- the second one is to use the glottal closure instants detection unit 808 , to generate GCI from the voice waveform.
- the glottal closure instants (GCI) 807 and the voice signal (PCM) 803 are sent to a processing unit 809 , to generate a complete set of segment points 810 .
- the voice signal in each frame 805 proceeds through a Fourier analysis unit 811 to generate amplitude spectrum 812 .
- the amplitude spectrum 812 proceeds through a Laguerre transform 813 to generate timbre vectors 814 .
- the timbre vectors 814 are streamed into acoustic decoder 815 , to compare with the timbre vectors stored in the acoustic models 816 .
- Possible phoneme sequence 817 is generated.
- the phoneme sequence is sent to language decoder 818 , assisted with language model 819 , to find the most probable output text 820 .
- the language decoder 818 may be essentially the same as other automatic speech recognition systems. Because the accuracy of the inventive parametric representation is much higher, the accuracy of the acoustic decoder 815 may be much higher.
- the PCM signals generated through a microphone can be sufficient.
- the addition of an electroglottograph 806 can substantially improve the accuracy.
Abstract
Description
- The present invention generally relates to voice transformation, in particular to voice transformation using orthogonal functions, and its applications in speech synthesis and automatic speech recognition.
- Voice transformation involves parameterization of a speech signal into a mathematical format which can be extensively manipulated such that the properties of the original speech, for example, pitch, speed, relative length of phones, prosody, and speaker identity, can be changed, but still sound natural. A straightforward application of voice transformation is singing synthesis. If the new parametric representation is successfully demonstrated to work well in voice transformation, it can be used for speech synthesis and automatic speech recognition.
- Speech synthesis, or text-to-speech (TTS), involves the use of a computer-based system to convert a written document into audible speech. A good TTS system should generate natural, or human-like, and highly intelligible speech. In the early years, the rule-based TTS systems, or the formant synthesizers, were used. These systems generate intelligible speech, but the speech sounds robotic, and unnatural.
- Currently, a great majority of commercial TTS systems are concatenative TTS system using the unit-selection method. According to this approach, a very large body of speech is recorded and stored. During the process of synthesis, the input text is first analyzed and the required prosodic features are predicted. Then, appropriate units are selected from a huge speech database, and stitched together. There are always mismatches at the border of consecutive segments from different origins. And there are always cases of required segments that do not exist in the speech database. Therefore, modifications of the recorder speech segments are necessary. Currently, the most popular method of speech modification is the time-domain pitch-synchronized overlap-add method (TD-PSOLA), LPC (linear prediction coefficients), mel-cepstral coefficients and sinusoidal representations. However, using those methods, the quality of voice is severely degraded. To improve the quality of speech synthesis and to allow for the use of a small database, voice transformation is the key. (See Part D of Springer Handbook of Speech Processing, Springer Verlag 2008).
- Automatic speech recognition (ASR) is the inverse process of speech synthesis. The first step, acoustic processing, reduces the speech signal into a parametric representation. Then, typically using HMM (Hidden Markov Model), with a statistic language model, the most likely text is thus produced. The state-of-the-art parametric representation for speech is LPC (linear prediction coefficients) and mel-cepstral coefficients. Obviously, the accuracy of speech parameterization affects the overall accuracy. (See Part E of Springer Handbook of Speech Processing, Springer Verlag 2008).
- The present invention is directed to a novel mathematical representation of the human voice as a timbre vector, together with a method of parameterizing speech into a timbre vector, and a method to recover human voice from a series of timbre vectors with variations. According to an exemplary embodiment of the invention, a speech signal is first segmented into non-overlapping frames using the glottal closure moment information. Using Fourier analysis, the speech signal in each frame is converted into amplitude spectrum, then Laguerre functions (based on a set of orthogonal polynomials) are used to convert the amplitude spectrum into a unit vector characteristic to the instantaneous timbre. A timbre vector is formed along with voicedness index, frame duration, and an intensity parameter. Because of the accuracy of the system and method and the complete separation of prosody and timbre, a variety of voice transformation operations can be applied, and the output voice is natural. A straightforward application of voice transformation is singing synthesis.
- One difference of the current invention from all previous methods is that the frames, or processing units, are non-overlapping, and do not require a window function. All previous parameterization methods, including linear prediction confidents, sinusoidal models, mel-cepstral coefficients and time-domain pitch synchronized overlap add methods rely on overlapping frames requiring a window function (such as Hamming window, Hann window, cosine window, triangular window, Gaussian window, etc.) and a shift time which is smaller than the duration of the frame, which makes an overlap.
- An important application of the inventive parametric representation is speech synthesis. Using the parametric representation in terms of timbre vectors, the speech segments can be modified to the prosodic requirements and regenerate an output speech with high quality. Furthermore, because of the complete separation of timbre and prosody data, the synthesized speech can have different speaker identity (baby, child, male, female, giant, etc), base pitch (up to three octaves), speed (up to 10 times), and various prosodic variations (calm, emotional, up to shouting). The timbre vector method disclosed in the present invention can be used to build high-quality speech synthesis systems using a compact speech database.
- Another important application of the inventive parametric representation of speech signal is to serve as the acoustic signal format to improve the accuracy of automatic speech recognition. The timbre vector method disclosed in the present invention can greatly improve the accuracy of automatic speech recognition.
-
FIG. 1 is a block diagram of a voice transformation systems using timbre vectors according to an exemplary embodiment of the present invention. -
FIG. 2 is an explanation of the basic concept of parameterization according to an exemplary embodiment of the present invention. -
FIG. 3 is the process of segmenting the PCM data according to an exemplary embodiment of the present invention. -
FIG. 4 is a plot of the Laguerre functions according to an exemplary embodiment of the present invention. -
FIG. 5 is the data structure of a timbre vector according to an exemplary embodiment of the present invention. -
FIG. 6 is the binomial interpolation of timbre vectors according to an exemplary embodiment of the present invention. -
FIG. 7 is a block diagram of a speech synthesis system using timbre vectors according to an exemplary embodiment of the present invention. -
FIG. 8 is a block diagram of an automatic speech recognition system using timbre vectors according to an exemplary embodiment of the present invention. - Various exemplary embodiments of the present invention are implemented on a computer system including one or more processors and one or more memory units. In this regard, according to exemplary embodiments, steps of the various methods described herein are performed on one or more computer processors according to instructions encoded on a computer-readable medium.
-
FIG. 1 is a block diagram of the voice transformation system according to an exemplary embodiment of the present invention. The source is the voice from aspeaker 101. Through amicrophone 102, the voice is converted into electrical signal, and recorded in the computer as PCM (Pulse Code Modulation)signal 103. ThePCM signal 103 is then segmented bysegmenter 104 intoframes 105, according tosegment points 110. There are two methods to generate the segment points. The first one is to use an electroglottograph (EGG) 106 to detect the glottal closure instants (GCI) 107 directly (SeeFIG. 2 ). The second one is to use a glottal closureinstants detection unit 108 to generate GCI from the voice waveform. The glottal closure instants (GCI) 107 and the voice signal (PCM) 103 are sent to aprocessing unit 109, to generate a complete set ofsegment points 110. The details of this process is shown inFIG. 3 . - The voice signal in each
frame 105 proceeds through a Fourieranalysis unit 111 to generateamplitude spectrum 112. Theamplitude spectrum 112 proceeds through anorthogonal transform unit 113 to generatetimbre vectors 114. In exemplary embodiments, Laguerre functions are the most appropriate mathematical functions for converting the amplitude spectrum into a compact and convenient form (seeFIG. 4 ). Data structure of a timbre vector is shown inFIG. 5 . - After the
PCM signal 103 is converted intotimbre vectors 114, a number of voice manipulations can be made according tospecifications 115 byvoice manipulator 116, so as to generatenew timbre vectors 117, then the voice can be regenerated using thenew timbre vectors 117. In detail, the steps are as follows: Laguerre transform 118 is used to regenerateamplitude spectrum 119; the phase generator 120 (based on Kramers-Kronig relations) is used to generatephase spectrum 121; FFT (Fast Fourier Transform) 122 is used to generate an elementaryacoustic wave 123, from the amplitude spectrum and phase spectrum; then those elementaryacoustic waves 123 are superposed according to thetiming information 124 in the new timbre vectors, each one is delayed by the time offrame duration 125 of the previous frame. The output wave in electric form then drives aloudspeaker 126 to produce anoutput voice 127. -
FIG. 2 shows the process of speech generation, particularly the generation of voiced sections, and the properties of the PCM and EGG signals.Air flow 201 comes from the lungs to the opening between the two vocal cords, or glottis, 202. If the glottis is constantly open, there is aconstant air flow 203, but no voice signal is generated. At the instant the glottis closes, or a glottal closure occurs, which is always very rapid due to the Bernoulli effect, the inertia of the moving air in thevocal track 204 generates a d'Alembert wave front, then excites an acoustic resonance. The actions of the glottis is monitored by the signals from a electroglottograph (EGG) 205. When there is a glottal closure, the instrument generates a sharp peak in the derivative of the EGG signal, as shown as 207 inFIG. 2 . Amicrophone 206 is placed near the mouth to generate a signal, typically a Pulse Code Modulation signal, or PCM, as shown in 209 inFIG. 2 . If the glottis remains closed after a closure, as shown as 208, then the acoustic excitation sustains, as shown as 210. -
FIG. 3 shows the details ofprocessing unit 109 to generate the segmentation points. The input data is the PCM signal 301-303 andEGG signal 304, produced by thesource speaker 101. When there are clear peaks in the EGG signal, such as 304, corresponding to PCM signal 301, those peaks are selected as the segmentation points 305. For some quasi-periodic segments of thevoice 302, there is no clear EGG peaks. The segmentation points are generated by comparing thewaveform 302 with the neighboringones 301, and if thewaveform 302 is still periodic, then segmentation points 306 are generated at the same intervals as the segmentation points 305. If the signal is no longer periodic, such as 303, the PCM is segmented according topoints 307 into frames with an equal interval, here 5 msec. Therefore, the entire PCM signal is segmented into frames. - The values of the voice signal at two adjacent closure moments may not match. The following is an algorithm that may be used to match the ends. Let the number of sampling points between two adjacent glottal closures be N, and the original voice signal be x0(n). The smoothed signal x(n) in a
small interval 0<n<M is defined as -
- Where M is about N/10. Otherwise x(n)=x0(n). Direct inspection shows that the ends of the waveform are matched, and it is smooth. Therefore, no window functions are required. The waveform in a frame is processed by Fourier analysis to generate an amplitude spectrum. The amplitude spectrum is further processed by a Laguarre transform unit to generate timbre vectors as follows.
- Laguerre functions are defined as
-
- where k is an integer, typically k=2 or k=4; and the associated Laguerre polynomials are
-
- The amplitude spectrum A(w) is expended into Laguerre functions
-
- where the coefficients are calculated by
-
- and κ is a scaling factor to maximize accuracy. The norm of the vector C is the intensity parameter I,
-
- and the normalized Laguarre coefficients are defined as
-
c n =C n /I. - To recover phase spectrum φ(Ω) from amplitude spectrum A(Ω), Kramers-Kronig relations are used,
-
- The output wave for a frame, the acoustic exciton, can be calculated from the amplitude spectrum A(Ω) and the phase spectrum φ(Ω),
-
-
FIG. 4 shows the Laguerre function. After proper scaling, twenty-nine Laguerre functions are used on thefrequency scale 401 of 0 to 11 kHz. Thefirst Laguerre function 402 actually probes the first formant. For higher order Laguerre functions, such as theLaguerre function 403, the resolution in the low-frequency range is successively improved; and extended to the high-frequency range 404. Because of the accuracy scaling, it makes an accurate but concise representation of the spectrum. -
FIG. 5 shows the data structure of a timbre vector including the voicedness index (V) 501, the frame duration (T) 502, the intensity parameter (I) 503, and the normalizedLaguerre coefficients 504. - There are many possible voice transformation manipulations, including, for example, the following:
- Timbre interpolation. The unit vector of Laguerre coefficients varies slowly with frames. It can be interpolated for reduced number of frames or extended number of frames for any section of voice to produce natural sounding speech of arbitrary temporal variations. For example, the speech can be made very fast but still recognizable by a blind person.
- Timbre fusing. By connecting two sets of timbre vectors of two different phonemes and smear-averaging over the juncture, a natural-sounding transition is generated. Phoneme assimilation may be automatically produced. By connecting a syllable ended with [g] with a syllable started with [n], after fusing, the sound [n] is automatically assimilated into [ng].
-
FIG. 6 shows the principles of the timbre fusing operation. Original timbre vectors from thefirst phoneme 601 include timbre vectors A, B, and C. Original timbre vectors from thesecond phoneme 602 include timbre vectors D and E. Theoutput timbre vectors 603 through 607 are weighed averages from the original timbre vectors. For example, output timbre vector D′ is generated from timbre vector C, D, and E using thebinomial coefficients binomial coefficients - Pitch modification. The state-of-the-art technology for pitch modification of speech signal is the time-domain pitch-synchronized overlap-add (TD-PSOLA) method, which can change pitch from −30% to +50%. Otherwise the output would sound unnatural. Here, pitch can be easily modified by changing the time of separation T, then using timbre interpolation to compensate speed. Natural sounding speech can be produced with pitch modifications as large as three octaves.
- Intensity profiling. Because the intensity parameter I is a property of a frame, it can be changed to produce any stress pattern required by prosody input.
- Change of speaker identity. First, by rescaling the amplitude spectrum on the frequency axis, the head size can be changed. The voice of an average adult speaker can be changed to that of a baby, a child, a woman, a man, or a giant. Second, by using a filter to alter the spectral envelop, special voice effects can be created.
- Using those voice manipulation capabilities and timbre fusing (see
FIG. 6 ), high-quality speech synthesizers with a compact database can be constructed using the parametric representation based on timbre vectors (seeFIG. 7 ). The speech synthesis system has two major parts: database building part 101 (the left-hand side ofFIG. 7 ), and the synthesis part 121 (right-hand side ofFIG. 7 ). - In the
database building unit 701, asource speaker 702 reads a prepared text. The voice is recorded by a microphone to become thePCM signal 703. The glottal closure signal is recorded by an electroglottograph (EGG) to becomeEGG signal 704. The origin and properties of those signals are shown inFIG. 2 . The EGG signal and the PCM signal are used by theprocessing unit 705 to generate a set of segment points 706. The details of the segmenting process, or the function of the processing unit, is shown inFIG. 3 . The PCM signal is segmented by thesegmenter 707 intoframes 708 using the segment points 706. Each frame is processed by a unit ofFourier analysis 709 to generateamplitude spectrum 710. The amplitude spectrum of each frame is then processed using aLaguerre transform unit 711 to become a unit vector, representing the instantaneous timbre of that frame, to become the basis oftimbre vectors 712. The Laguerre functions are shown inFIG. 4 . The structure of the timbre vector is shown in FIG. 5. The timbre vectors of various units of speech, such as, for example, phonemes, diphones, demisyllables, syllables, words and even phrases, are then stored in thespeech database 720. - In the
synthesis unit 721, theinput text 722 together withsynthesis parameters 723, are fed into thefrontend 724. Detailed instructions about the phonemes, intensity andpitch values 725, for generating the desired speech are generated, then input to aprocessing unit 726. Theprocessing unit 726 selects timbre vectors from thedatabase 720, then converts the selected timbre vectors to a new series oftimbre vectors 727 according to the instructions from theprocess unit 726, and using timbre fusing if necessary (seeFIG. 6 ). Each timbre vector is converted into anamplitude spectrum 729 byLaguerre transform unit 728. Thephase spectrum 731 is generated from theamplitude spectrum 729 byphase generator 730 using a Kramers-Kronig relations algorithm. Theamplitude spectrum 729 and thephase spectrum 731 are sent to a FFT (Fast Fourier Transform)unit 732, to generate an elementaryacoustic wave 733. Those elementaryacoustic waves 733 are than superposed by thesuperposition unit 735 according to thetiming information 734 provided by thenew timbre vectors 727, to generate the final result,output speech signal 736. - The parametric representation of human voice in terms of timbre vectors can also be used as the basis of automatic speech recognition systems. To date, the most widely used acoustic features, or parametric representation of human speech in automatic speech recognition is the mel-cepstrum. First, the speech signal is segmented into frames of fixed length, typically 20 msec, with a window, typically Hann window or Hamming window, and a shift of 10 msec. Those parametric representations are crude and inaccurate. Features that cross the phoneme borders occur very often.
- The parametric representation based on timbre vectors is more accurate. Especially, a well-behaved timbre distance δ between two frames can be defined as
-
- where c(1) n and c(2) n are elements of the normalized Laguerre coefficients of the two timbre vectors (see
FIG. 5 ). Experiments have shown that for two timbre vectors of the same phoneme (not diphthong), the distance is less than 0.1. For timbre vectors of different vowels, the distance is 0.1 to 0.6. Furthermore, because of the presence of the voicedness index V (seeFIG. 5 ), vowels and unvoiced consonants are well separated. Because of the intensity parameter I, silence is well separated from real sound. For the recognition of tone languages such as Mandarin, Cantonese, Thai etc., pitch is an important parameter (see, for example, U.S. Pat. No. 5,751,905 and U.S. Pat. No. 6,510,410). The frame duration T provides a very accurate measure of pitch (seeFIG. 5 ). Therefore, using parametric representation based on timbre vectors, the accuracy of speech recognition can be greatly improved. -
FIG. 8 shows a block diagram of an automatic speech recognition system based on timbre vectors. The first half of the procedure, converting speech signal into timbre vectors, is similar to step 102 throughstep 114 ofFIG. 1 for voice transformation. The voice from aspeaker 801 is recorded in the computer asPCM signal 803. ThePCM signal 803 is then segmented bysegmenter 804 intoframes 805, according to segment points 810. There are two methods to generate the segment points. The first one is to use an electroglottograph (EGG) 806 to detect the glottal closure instants (GCI) 807 directly (seeFIG. 2 ). The second one is to use the glottal closureinstants detection unit 808, to generate GCI from the voice waveform. The glottal closure instants (GCI) 807 and the voice signal (PCM) 803 are sent to aprocessing unit 809, to generate a complete set of segment points 810. The details of this process are shown inFIG. 3 . - The voice signal in each
frame 805 proceeds through aFourier analysis unit 811 to generateamplitude spectrum 812. Theamplitude spectrum 812 proceeds through aLaguerre transform 813 to generatetimbre vectors 814. - The
timbre vectors 814 are streamed intoacoustic decoder 815, to compare with the timbre vectors stored in theacoustic models 816.Possible phoneme sequence 817 is generated. The phoneme sequence is sent tolanguage decoder 818, assisted withlanguage model 819, to find the mostprobable output text 820. Thelanguage decoder 818 may be essentially the same as other automatic speech recognition systems. Because the accuracy of the inventive parametric representation is much higher, the accuracy of theacoustic decoder 815 may be much higher. - For using the speech recognition system in a quiet environment, the PCM signals generated through a microphone can be sufficient. In noisy environments, the addition of an
electroglottograph 806 can substantially improve the accuracy. - In ordinary speech recognition systems, adaptation for a given speaker by recording a good number (for example 100) of spoken sentences from a given speaker and processing it can improve the accuracy. Because of the simplicity of the timbre-vector parametric representation, it is possible to use a single recorded sentence from a given speaker to improve the accuracy.
- While this invention has been described in conjunction with the exemplary embodiments outlined above, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, the exemplary embodiments of the invention, as set forth above, are intended to be illustrative, not limiting. Various changes may be made without departing from the spirit and scope of the invention.
Claims (20)
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/625,317 US8744854B1 (en) | 2012-09-24 | 2012-09-24 | System and method for voice transformation |
US13/692,584 US8719030B2 (en) | 2012-09-24 | 2012-12-03 | System and method for speech synthesis |
US13/692,621 US20140088968A1 (en) | 2012-09-24 | 2012-12-03 | System and method for speech recognition using timbre vectors |
PCT/US2013/053549 WO2014046789A1 (en) | 2012-09-24 | 2013-08-05 | System and method for voice transformation, speech synthesis, and speech recognition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/625,317 US8744854B1 (en) | 2012-09-24 | 2012-09-24 | System and method for voice transformation |
Related Child Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US13/692,621 Continuation US20140088968A1 (en) | 2012-09-24 | 2012-12-03 | System and method for speech recognition using timbre vectors |
US13/692,584 Continuation US8719030B2 (en) | 2012-09-24 | 2012-12-03 | System and method for speech synthesis |
Publications (2)
Publication Number | Publication Date |
---|---|
US20140142946A1 true US20140142946A1 (en) | 2014-05-22 |
US8744854B1 US8744854B1 (en) | 2014-06-03 |
Family
ID=50339719
Family Applications (3)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US13/625,317 Active US8744854B1 (en) | 2012-09-24 | 2012-09-24 | System and method for voice transformation |
US13/692,584 Active US8719030B2 (en) | 2012-09-24 | 2012-12-03 | System and method for speech synthesis |
US13/692,621 Abandoned US20140088968A1 (en) | 2012-09-24 | 2012-12-03 | System and method for speech recognition using timbre vectors |
Family Applications After (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US13/692,584 Active US8719030B2 (en) | 2012-09-24 | 2012-12-03 | System and method for speech synthesis |
US13/692,621 Abandoned US20140088968A1 (en) | 2012-09-24 | 2012-12-03 | System and method for speech recognition using timbre vectors |
Country Status (2)
Country | Link |
---|---|
US (3) | US8744854B1 (en) |
WO (1) | WO2014046789A1 (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140130655A1 (en) * | 2012-11-13 | 2014-05-15 | Yamaha Corporation | Delayed registration data readout in electronic music apparatus |
US20140156280A1 (en) * | 2012-11-30 | 2014-06-05 | Kabushiki Kaisha Toshiba | Speech processing system |
WO2015183254A1 (en) * | 2014-05-28 | 2015-12-03 | Interactive Intelligence, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
CN106920547A (en) * | 2017-02-21 | 2017-07-04 | 腾讯科技(上海)有限公司 | Phonetics transfer method and device |
US10014007B2 (en) | 2014-05-28 | 2018-07-03 | Interactive Intelligence, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
US10255903B2 (en) | 2014-05-28 | 2019-04-09 | Interactive Intelligence Group, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
CN111179905A (en) * | 2020-01-10 | 2020-05-19 | 北京中科深智科技有限公司 | Rapid dubbing generation method and device |
US10978076B2 (en) * | 2017-03-22 | 2021-04-13 | Kabushiki Kaisha Toshiba | Speaker retrieval device, speaker retrieval method, and computer program product |
CN112820267A (en) * | 2021-01-15 | 2021-05-18 | 科大讯飞股份有限公司 | Waveform generation method, training method of related model, related equipment and device |
CN113066472A (en) * | 2019-12-13 | 2021-07-02 | 科大讯飞股份有限公司 | Synthetic speech processing method and related device |
Families Citing this family (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
FR3020732A1 (en) * | 2014-04-30 | 2015-11-06 | Orange | PERFECTED FRAME LOSS CORRECTION WITH VOICE INFORMATION |
CN104021148B (en) * | 2014-05-16 | 2017-05-03 | 小米科技有限责任公司 | Method and device for adjusting sound effect |
US9607610B2 (en) * | 2014-07-03 | 2017-03-28 | Google Inc. | Devices and methods for noise modulation in a universal vocoder synthesizer |
US9685169B2 (en) * | 2015-04-15 | 2017-06-20 | International Business Machines Corporation | Coherent pitch and intensity modification of speech signals |
JP6496030B2 (en) * | 2015-09-16 | 2019-04-03 | 株式会社東芝 | Audio processing apparatus, audio processing method, and audio processing program |
US10861476B2 (en) * | 2017-05-24 | 2020-12-08 | Modulate, Inc. | System and method for building a voice database |
CN107240401B (en) * | 2017-06-13 | 2020-05-15 | 厦门美图之家科技有限公司 | Tone conversion method and computing device |
JP7147211B2 (en) * | 2018-03-22 | 2022-10-05 | ヤマハ株式会社 | Information processing method and information processing device |
CN108830232B (en) * | 2018-06-21 | 2021-06-15 | 浙江中点人工智能科技有限公司 | Voice signal period segmentation method based on multi-scale nonlinear energy operator |
CN111048062B (en) * | 2018-10-10 | 2022-10-04 | 华为技术有限公司 | Speech synthesis method and apparatus |
CN110138654B (en) * | 2019-06-06 | 2022-02-11 | 北京百度网讯科技有限公司 | Method and apparatus for processing speech |
WO2021030759A1 (en) | 2019-08-14 | 2021-02-18 | Modulate, Inc. | Generation and detection of watermark for real-time voice conversion |
CN110689885B (en) * | 2019-09-18 | 2023-05-23 | 平安科技(深圳)有限公司 | Machine synthesized voice recognition method, device, storage medium and electronic equipment |
CN110808026B (en) * | 2019-11-04 | 2022-08-23 | 金华航大北斗应用技术有限公司 | Electroglottography voice conversion method based on LSTM |
CN111435591B (en) * | 2020-01-17 | 2023-06-20 | 珠海市杰理科技股份有限公司 | Voice synthesis method and system, audio processing chip and electronic equipment |
CN111489734B (en) * | 2020-04-03 | 2023-08-22 | 支付宝(杭州)信息技术有限公司 | Model training method and device based on multiple speakers |
CN114203147A (en) | 2020-08-28 | 2022-03-18 | 微软技术许可有限责任公司 | System and method for text-to-speech cross-speaker style delivery and for training data generation |
CN112562728A (en) * | 2020-11-13 | 2021-03-26 | 百果园技术(新加坡)有限公司 | Training method for generating confrontation network, and audio style migration method and device |
CN112365882B (en) * | 2020-11-30 | 2023-09-22 | 北京百度网讯科技有限公司 | Speech synthesis method, model training method, device, equipment and storage medium |
CN112562635B (en) * | 2020-12-03 | 2024-04-09 | 云知声智能科技股份有限公司 | Method, device and system for solving generation of pulse signals at splicing position in speech synthesis |
CN112652318B (en) * | 2020-12-21 | 2024-03-29 | 北京捷通华声科技股份有限公司 | Tone color conversion method and device and electronic equipment |
KR102576606B1 (en) * | 2021-03-26 | 2023-09-08 | 주식회사 엔씨소프트 | Apparatus and method for timbre embedding model learning |
RU2763124C1 (en) * | 2021-07-06 | 2021-12-27 | Валерий Олегович Лелейтнер | Method for speaker-independent phoneme recognition in a speech signal |
CN116246643B (en) * | 2022-12-26 | 2023-07-28 | 深度好奇(杭州)科技有限公司 | Speech robot and artificial seat tone normalization method, system and equipment |
Family Cites Families (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
FR2636163B1 (en) | 1988-09-02 | 1991-07-05 | Hamon Christian | METHOD AND DEVICE FOR SYNTHESIZING SPEECH BY ADDING-COVERING WAVEFORMS |
DE69231266T2 (en) | 1991-08-09 | 2001-03-15 | Koninkl Philips Electronics Nv | Method and device for manipulating the duration of a physical audio signal and a storage medium containing such a physical audio signal |
KR940002854B1 (en) | 1991-11-06 | 1994-04-04 | 한국전기통신공사 | Sound synthesizing system |
US5751905A (en) | 1995-03-15 | 1998-05-12 | International Business Machines Corporation | Statistical acoustic processing method and apparatus for speech recognition using a toned phoneme system |
DE69612958T2 (en) | 1995-11-22 | 2001-11-29 | Koninkl Philips Electronics Nv | METHOD AND DEVICE FOR RESYNTHETIZING A VOICE SIGNAL |
GB9600774D0 (en) * | 1996-01-15 | 1996-03-20 | British Telecomm | Waveform synthesis |
US5905972A (en) * | 1996-09-30 | 1999-05-18 | Microsoft Corporation | Prosodic databases holding fundamental frequency templates for use in speech synthesis |
KR100269255B1 (en) | 1997-11-28 | 2000-10-16 | 정선종 | Pitch Correction Method by Variation of Gender Closure Signal in Voiced Signal |
CA2296330C (en) * | 1997-07-31 | 2009-07-21 | British Telecommunications Public Limited Company | Generation of voice messages |
US6304846B1 (en) * | 1997-10-22 | 2001-10-16 | Texas Instruments Incorporated | Singing voice synthesis |
US6101470A (en) * | 1998-05-26 | 2000-08-08 | International Business Machines Corporation | Methods for generating pitch and duration contours in a text to speech system |
US20070233479A1 (en) * | 2002-05-30 | 2007-10-04 | Burnett Gregory C | Detecting voiced and unvoiced speech using both acoustic and nonacoustic sensors |
US6510410B1 (en) | 2000-07-28 | 2003-01-21 | International Business Machines Corporation | Method and apparatus for recognizing tone languages using pitch information |
US20020099541A1 (en) * | 2000-11-21 | 2002-07-25 | Burnett Gregory C. | Method and apparatus for voiced speech excitation function determination and non-acoustic assisted feature extraction |
WO2004068467A1 (en) * | 2003-01-31 | 2004-08-12 | Oticon A/S | Sound system improving speech intelligibility |
SG120121A1 (en) * | 2003-09-26 | 2006-03-28 | St Microelectronics Asia | Pitch detection of speech signals |
US20070011009A1 (en) | 2005-07-08 | 2007-01-11 | Nokia Corporation | Supporting a concatenative text-to-speech synthesis |
JP4129989B2 (en) * | 2006-08-21 | 2008-08-06 | インターナショナル・ビジネス・マシーンズ・コーポレーション | A system to support text-to-speech synthesis |
EP1970894A1 (en) * | 2007-03-12 | 2008-09-17 | France Télécom | Method and device for modifying an audio signal |
CN101281744B (en) * | 2007-04-04 | 2011-07-06 | 纽昂斯通讯公司 | Method and apparatus for analyzing and synthesizing voice |
US8401849B2 (en) * | 2008-12-18 | 2013-03-19 | Lessac Technologies, Inc. | Methods employing phase state analysis for use in speech synthesis and recognition |
JP5433696B2 (en) * | 2009-07-31 | 2014-03-05 | 株式会社東芝 | Audio processing device |
WO2011068417A1 (en) * | 2009-12-01 | 2011-06-09 | Jevon Joseph Longdell | Method and apparatus for detection of ultrasound |
-
2012
- 2012-09-24 US US13/625,317 patent/US8744854B1/en active Active
- 2012-12-03 US US13/692,584 patent/US8719030B2/en active Active
- 2012-12-03 US US13/692,621 patent/US20140088968A1/en not_active Abandoned
-
2013
- 2013-08-05 WO PCT/US2013/053549 patent/WO2014046789A1/en active Application Filing
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9111514B2 (en) * | 2012-11-13 | 2015-08-18 | Yamaha Corporation | Delayed registration data readout in electronic music apparatus |
US20140130655A1 (en) * | 2012-11-13 | 2014-05-15 | Yamaha Corporation | Delayed registration data readout in electronic music apparatus |
US9466285B2 (en) * | 2012-11-30 | 2016-10-11 | Kabushiki Kaisha Toshiba | Speech processing system |
US20140156280A1 (en) * | 2012-11-30 | 2014-06-05 | Kabushiki Kaisha Toshiba | Speech processing system |
US10014007B2 (en) | 2014-05-28 | 2018-07-03 | Interactive Intelligence, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
WO2015183254A1 (en) * | 2014-05-28 | 2015-12-03 | Interactive Intelligence, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
US10255903B2 (en) | 2014-05-28 | 2019-04-09 | Interactive Intelligence Group, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
US10621969B2 (en) | 2014-05-28 | 2020-04-14 | Genesys Telecommunications Laboratories, Inc. | Method for forming the excitation signal for a glottal pulse model based parametric speech synthesis system |
CN106920547A (en) * | 2017-02-21 | 2017-07-04 | 腾讯科技(上海)有限公司 | Phonetics transfer method and device |
US10978076B2 (en) * | 2017-03-22 | 2021-04-13 | Kabushiki Kaisha Toshiba | Speaker retrieval device, speaker retrieval method, and computer program product |
CN113066472A (en) * | 2019-12-13 | 2021-07-02 | 科大讯飞股份有限公司 | Synthetic speech processing method and related device |
CN111179905A (en) * | 2020-01-10 | 2020-05-19 | 北京中科深智科技有限公司 | Rapid dubbing generation method and device |
CN112820267A (en) * | 2021-01-15 | 2021-05-18 | 科大讯飞股份有限公司 | Waveform generation method, training method of related model, related equipment and device |
Also Published As
Publication number | Publication date |
---|---|
US8744854B1 (en) | 2014-06-03 |
US20140088958A1 (en) | 2014-03-27 |
US8719030B2 (en) | 2014-05-06 |
WO2014046789A1 (en) | 2014-03-27 |
US20140088968A1 (en) | 2014-03-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8744854B1 (en) | System and method for voice transformation | |
Drugman et al. | Glottal source processing: From analysis to applications | |
Toda et al. | Statistical mapping between articulatory movements and acoustic spectrum using a Gaussian mixture model | |
US6804649B2 (en) | Expressivity of voice synthesis by emphasizing source signal features | |
JP4355772B2 (en) | Force conversion device, speech conversion device, speech synthesis device, speech conversion method, speech synthesis method, and program | |
JP3408477B2 (en) | Semisyllable-coupled formant-based speech synthesizer with independent crossfading in filter parameters and source domain | |
Degottex et al. | Mixed source model and its adapted vocal tract filter estimate for voice transformation and synthesis | |
Bellegarda et al. | Statistical prosodic modeling: from corpus design to parameter estimation | |
Reddy et al. | Excitation modelling using epoch features for statistical parametric speech synthesis | |
Lee et al. | Spectral and prosodic transformations of hearing-impaired Mandarin speech | |
Mann | An investigation of nonlinear speech synthesis and pitch modification techniques | |
Fahad et al. | Synthesis of emotional speech by prosody modification of vowel segments of neutral speech | |
Wang et al. | Emotional voice conversion for mandarin using tone nucleus model–small corpus and high efficiency | |
Waghmare et al. | Analysis of pitch and duration in speech synthesis using PSOLA | |
Del Pozo | Voice source and duration modelling for voice conversion and speech repair | |
Csapó et al. | A novel irregular voice model for HMM-based speech synthesis. | |
Matsuda et al. | Applying generation process model constraint to fundamental frequency contours generated by hidden-Markov-model-based speech synthesis | |
Reddy et al. | Neutral to joyous happy emotion conversion | |
Lehana et al. | Improving quality of speech synthesis in Indian Languages | |
Dutoit et al. | Speech Synthesis | |
Anil et al. | Pitch and duration modification for expressive speech synthesis in Marathi TTS system | |
Lavner et al. | Voice morphing using 3D waveform interpolation surfaces and lossless tube area functions | |
Li | Prosody analysis and modeling for Cantonese text-to-speech | |
Nguyen | Studies on spectral modification in voice transformation | |
Ferencz et al. | THE NEW VERSION OF THE ROMVOX TEXT-TO-SPEECH SYNTHESIS SYSTEM BASED ON A HYBRID TIME DOMAIN-LPC SYNTHESIS |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
AS | Assignment |
Owner name: THE TRUSTEES OF COLUMBIA UNIVERSITY IN THE CITY OF Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:CHEN, CHENGJUN JULIAN;REEL/FRAME:037522/0331 Effective date: 20160114 |
|
MAFP | Maintenance fee payment |
Free format text: PAYMENT OF MAINTENANCE FEE, 4TH YR, SMALL ENTITY (ORIGINAL EVENT CODE: M2551) Year of fee payment: 4 |
|
MAFP | Maintenance fee payment |
Free format text: PAYMENT OF MAINTENANCE FEE, 8TH YR, SMALL ENTITY (ORIGINAL EVENT CODE: M2552); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY Year of fee payment: 8 |