US5794192A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Self-learning speaker adaptation based on spectral bias source decomposition, using very short calibration speech
Assignee: Panasonic Corp of North America
•
Filing: 1996-09-12
•
Granted: 1998-08-11
•
Expires: 2016-09-12
Abstract
Abstract A speaker adaptation technique based on the separation of speech spectra variation sources is developed for improving speaker-independent continuous speech recognition. The variation sources include speaker acoustic characteristics, and contextual dependency of allophones. Statistical methods are formulated to normalize speech spectra based on speaker acoustic characteristics and then adapt mixture Gaussian density phone models based on speaker phonologic characteristics. Adaptation experiments using short calibration speech (5 sec./speaker) have shown substantial performance improvement over the baseline recognition system.
Inventors
- 👤 Yunxin Zhao
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 15:20
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/10; G10L15/06; G10L15/063; G10L15/065; G10L15/07; G10L15/20; G10L21/00; G10L21/02
Claims Analysis
Total Claims: 17
Claim 1
Independent Claim
315 words
A speech recognition method comprising the steps of: a. providing training speech that includes a passage of calibration speech for each training speaker; b. representing the training speech in a spectral domain such that each training speech utterance is represented by a sequence of training speech spectra; c. building a first set of Gaussian density phone models from the spectra of all calibration speech; d. estimating a spectral bias indicative of speaker acoustic characteristics for each calibration speech using said first set of Gaussian density phone models; e. normalizing the training speech spectra based on speaker acoustic characteristics using said spectral bias; f. building a second set of Gaussian mixture density phone models having parameters of mean vectors, covariance matrices and mixture weights from said normalized training speech spectra; g. taking a passage of calibration speech from each speaker; h. representing the calibration speech in a spectral domain such that each calibration speech utterance is represented by a sequence of speech spectra; i. estimating a spectral bias indicative of speaker acoustic characteristics for each calibration speech using said second set of Gaussian mixture density phone models built in step f; j. normalizing the calibration speech spectra based on speaker acoustic characteristics using said spectral bias; k. adapting the phone model parameters based on speaker phonologic characteristics using the normalized calibration speech, where context modulation vectors are estimated between Gaussian densities in each mixture, and the context modulation vectors are used to shift the spectra of the calibration speech; l. providing test speech for speech recognition; m. representing the test speech in a spectral domain such that the test speech is represented by a sequence of test speech spectra; n. normalizing the test speech spectra based on speaker acoustic characteristics using said spectral bias; o. using the normalized test speech spectra in conjunction with the adapted Gaussian mixture density phone models to recognize the test speech.
Claim 2
Dependent Claim
315 words
A speech recognition method comprising the steps of: a. providing training speech that includes a passage of calibration speech for each training speaker; b. representing the training speech in a spectral domain such that each training speech utterance is represented by a sequence of training speech spectra; c. building a first set of Gaussian density phone models from the spectra of all calibration speech; d. estimating a spectral bias indicative of speaker acoustic characteristics for each calibration speech using said first set of Gaussian density phone models; e. normalizing the training speech spectra based on speaker acoustic characteristics using said spectral bias; f. building a second set of Gaussian mixture density phone models having parameters of mean vectors, covariance matrices and mixture weights from said normalized training speech spectra; g. taking a passage of calibration speech from each speaker; h. representing the calibration speech in a spectral domain such that each calibration speech utterance is represented by a sequence of speech spectra; i. estimating a spectral bias indicative of speaker acoustic characteristics for each calibration speech using said second set of Gaussian mixture density phone models built in step f; j. normalizing the calibration speech spectra based on speaker acoustic characteristics using said spectral bias; k. adapting the phone model parameters based on speaker phonologic characteristics using the normalized calibration speech, where context modulation vectors are estimated between Gaussian densities in each mixture, and the context modulation vectors are used to shift the spectra of the calibration speech; l. providing test speech for speech recognition; m. representing the test speech in a spectral domain such that the test speech is represented by a sequence of test speech spectra; n. normalizing the test speech spectra based on speaker acoustic characteristics using said spectral bias; o. using the normalized test speech spectra in conjunction with the adapted Gaussian mixture density phone models to recognize the test speech.
Claim 3
Dependent Claim
35 words
The method of claim 1 wherein the step of providing training speech is performed by providing sample speech from a plurality of persons that includes calibration speech consisting of the same predefined set of words.
Claim 4
Dependent Claim
17 words
The method of claim 2 wherein the predefined set of words is uttered in continuous speech fashion.
Claim 5
Dependent Claim
29 words
The method of claim 1 wherein said step of representing the training speech in a spectral domain comprises extracting PLP cepstrum coefficients indicative of phonetic features of the speech.
Claim 6
Dependent Claim
30 words
The method of claim 1 wherein said step of representing the training speech in a spectral domain comprises extracting first-order temporal regression coefficients to represent dynamic features of the speech.
Claim 7
Dependent Claim
26 words
The method of claim 1 wherein said normalizing steps are performed by estimating the spectral deviation vector and subsequently removing said vector from the speech spectra.
Claim 8
Dependent Claim
26 words
The method of claim 1 wherein the step of normalizing the training speech spectra is performed by estimating the parameters of unimodel Gaussian density phone models.
Claim 9
Dependent Claim
36 words
The method of claim 7 further comprising using said phone models to estimate the spectral deviation vector for each of the speakers and subsequently removing said vector from the speech spectra for each of the speakers.
Claim 10
Dependent Claim
69 words
The method of claim 1 wherein the step of normalizing the training speech spectra is performed by: (1) generating a set of unimodal Gaussian density phone models from the calibration speech; and (2) using said set of unimodal Gaussian density phone models to estimate the spectral deviation vector for each of the speakers and subsequently removing said spectral deviation vector from the speech spectra for each of the speakers.
Claim 11
Dependent Claim
34 words
The method of claim 1 wherein the step of adapting the phone model parameters is performed by modifying the Gaussian mixture density parameters based on context-modulated acoustically normalized calibration speech from a specific speaker.
Claim 12
Dependent Claim
25 words
The method of claim 10 wherein the context-modulated calibration speech is generated by subtracting context modulation vectors from the calibration speech of said specific speaker.
Claim 13
Dependent Claim
29 words
The method of claim 11 wherein said context modulation vectors are estimated based on training data from a plurality of training speakers and said Gaussian mixture density phone model.
Claim 14
Dependent Claim
45 words
The method of claim 11 wherein said subtracting comprises subtracting a context modulation vector from a segment of a phone unit in the calibration speech of said specific speaker for each Gaussian density in the Gaussian mixture density for the state of the phone unit.
Claim 15
Dependent Claim
24 words
The method of claim 13 wherein the phone unit segment is obtained by an automatic segmentation of the calibration speech of the specific speaker.
Claim 16
Dependent Claim
23 words
The method of claim 13 wherein the phone unit segment is obtained by Viterbi segmentation of the calibration speech of the specific speaker.
Claim 17
Dependent Claim
20 words
The method of claim 10 wherein said step of modifying the Gaussian mixture density parameters is performed by Bayesian estimation.
References Cited (Backward Citations)
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2001-08-10 | 2002-02-09 | 2002-08-11 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2005-08-10 | 2006-02-09 | 2006-08-11 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2009-08-10 | 2010-02-09 | 2010-08-11 | $7700.00 | Window Closed / Lapsed |