US5721808A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Method for the composition of noise-resistant hidden markov models for speech recognition and speech recognizer using the same
Assignee: NTT Inc
•
Filing: 1996-03-04
•
Granted: 1998-02-24
•
Expires: 2016-03-04
Abstract
Abstract Noise-resistant speech HMMs are composed by: recording noise in the environment of utterance (S1); preparing HMMs of the noise (S2); transforming the output probability distribution of each of the noise HMMs and speech HMMs prepared from speech unaffected by noise and multiplicative distortion to a linear spectral domain (S31); multiplying the speech HMM distribution in the linear spectral domain by a multiplicative distortion W that is an unknown variable (S321); convoluting the multiplied value and the noise HMM distribution in the linear spectral domain (S322); inversely transforming the convoluted value to the original domain of the speech HMM (S33) to compose incomplete noise-resistant speech HMMs each containing multiplicative distortion as an unknown variable (S3); calculating the likelihoods of the incomplete noise-resistant speech HMMs for input speech and estimating the multiplicative distortion of that one of the incomplete noise-resistant HMMs which has the maximum likelihood (S4); and substituting the estimated value into the incomplete noise-resistant speech HMMs (S5).
Inventors
- 👤 Yasuhiro Minami
- 👤 Tomoko Matsui
- 👤 Sadaoki Furui
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-07 18:06
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/20; G10L15/08; G10L15/14; G10L15/142; G10L15/144
Claims Analysis
Total Claims: 23
Claim 1
Independent Claim
135 words
A method for the composition of noise-resistant Hidden Markov Models (HMMs) for speech recognition, comprising: a first step of preparing, from speech HMMs which have been produced from speech information free of noise or multiplicative distortion and noise HMMs which have been produced from environmental noise, incomplete noise-resistant speech HMMs each containing said multiplicative distortion or an SN (Signal-to-Noise) ratio in a linear spectral domain as an unknown variable; a second step of estimating said multiplicative distortion or SN ratio that maximizes the likelihood of each of said incomplete noise-resistant speech HMMs for input speech; and a third step of substituting said estimated values into said incomplete noise-resistant speech HMMs to complete said noise-resistant speech HMMs and storing said noise-resistant speech HMMs in a storage means for subsequent use in the recognition of input speech.
Claim 2
Independent Claim
135 words
A method for the composition of noise-resistant Hidden Markov Models (HMMs) for speech recognition, comprising: a first step of preparing, from speech HMMs which have been produced from speech information free of noise or multiplicative distortion and noise HMMs which have been produced from environmental noise, incomplete noise-resistant speech HMMs each containing said multiplicative distortion or an SN (Signal-to-Noise) ratio in a linear spectral domain as an unknown variable; a second step of estimating said multiplicative distortion or SN ratio that maximizes the likelihood of each of said incomplete noise-resistant speech HMMs for input speech; and a third step of substituting said estimated values into said incomplete noise-resistant speech HMMs to complete said noise-resistant speech HMMs and storing said noise-resistant speech HMMs in a storage means for subsequent use in the recognition of input speech.
Claim 3
Dependent Claim
49 words
The method of claim 1, wherein said first step comprises: a fourth step of adding said multiplicative distortion to each of said speech HMMs; and a fifth step of combining said each speech HMM added with said multiplicative distortion and each of said noise HMMs in a product space.
Claim 4
Dependent Claim
35 words
The method of claim 2, wherein said fourth step is a step of adding said multiplicative distortion to a mean value of the output probability distribution of said speech HMMs represented in a cepstral domain.
Claim 5
Dependent Claim
36 words
The method of claim 2, wherein said fourth step is a step of adding said multiplicative distortion to a means value of the output probability distribution of said speech HMMs represented in a logarithm spectral domain.
Claim 6
Dependent Claim
51 words
The method of claim 1, wherein said first step comprises: a fourth step of adding one of said speech and said noise HMMs with a component containing said SN ratio; and a fifth step of combining said one HMM added with said component and the other HMM in a product space.
Claim 7
Dependent Claim
42 words
The method of claim 5, wherein said fourth step is a step of adding said component containing said SN ratio to the 0-order term of a mean value of the output probability distribution of said speech HMMs represented in a cepstrum domain.
Claim 8
Dependent Claim
39 words
The method of claim 5, wherein said fourth step is a step of adding said component containing said SN ratio to a mean value of the output probability distribution of said speech HMMs represented in a logarithm spectral domain.
Claim 9
Dependent Claim
102 words
The method of claim 1, wherein said first step comprises: a sixth step of transforming the output probability distributions of said speech and noise HMMs to said linear spectral domain; a seventh step of multiplying said output probability distribution of said speech HMMs in said linear spectral domain by said multiplicative distortion; an eighth step of convoluting said output probability distribution of said speech HMMs multiplied by said multiplicative distortion and said output probability distribution of said noise HMMs in said linear spectral domain; and a ninth step of inversely transforming said convoluted value to the original domain of said speech HMMs.
Claim 10
Dependent Claim
103 words
The method of claim 1, wherein said first step comprises: a sixth step of transforming the output probability distributions of said speech and noise HMMs to said linear spectral domain; a seventh step of multiplying said output probability distribution of one of said speech and noise HMMs in said linear spectral domain by a component containing said SN ratio; an eighth step of convoluting said output probability distribution of said one HMM multiplied by said component and said output probability distribution of the other HMM; and a ninth step of inversely transforming said convoluted value to the original domain of said speech HMMs.
Claim 11
Dependent Claim
64 words
The method of claim 8 or 9, wherein said sixth step is a step of cosine-transforming said output probability distributions of said speech and noise HMMs represented in said cepstral domain and exponentially transforming said cosine-transformed output probability distributions of said speech and noise HMMs and said ninth step is a step of logarithm-transforming said convoluted value and inversely cosine-transforming said logarithm-transformed convoluted value.
Claim 12
Dependent Claim
45 words
The method of claim 8 or 9, wherein said sixth step is a step of exponentially transforming said output probability distributions of said speech and noise HMMs represented in said logarithm spectral domain and said ninth step is a step of logarithm-transforming said convoluted value.
Claim 13
Dependent Claim
62 words
The method of claim 1, wherein said first step comprises: a fourth step of multiplying the output probability distribution of said speech HMMs represented in said linear spectral domain by said multiplicative distortion; and a step of convoluting said output probability distribution multiplied by said multiplicative distortion and the output probability distribution of said noise HMMs represented in said linear spectral domain.
Claim 14
Dependent Claim
67 words
The method of claim 1, wherein said first step comprises: a fourth step of multiplying the output probability distribution of one of said speech and noise HMMs represented in said linear spectral domain by a component containing said SN ratio; and a fifth step of convoluting said output probability distribution of said one HMM multiplied by said component and said output probability distribution of the other HMM.
Claim 15
Dependent Claim
17 words
The method of claim 5, 9, or 13, wherein said component containing said SN ratio is 10.sup.(S/N)/2).
Claim 16
Independent Claim
35 words
The method of any one of claims 1 to 9, 12 or 13, wherein said second step is a step of making said estimation by a repetitive operation based on a maximum likelihood estimation method.
Claim 17
Independent Claim
34 words
The method of any one of claims 1 to 9, 12 or 13, wherein said second step is a step of making said estimation by a repetitive operation based on a steepest descent method.
Claim 18
Independent Claim
71 words
The method of any one of claims 1 to 4, 8 or 12, wherein said second step is a step of multiplying each of said incomplete noise-resistant speech HMMs by a different multiplicative distortion, calculating the likelihood of said each incomplete noise-resistant speech HMM for said input speech, and estimating the value of the multiplicative distortion corresponding to that one of said incomplete noise-resistant speech HMMs which has the maximum likelihood.
Claim 19
Independent Claim
71 words
The method of any one of claims 1, 5 to 7, 9 or 13, wherein said second step is a step of multiplying each of said incomplete noise-resistant speech HMMs by a different SN ratio, calculating the likelihood of said each incomplete noise-resistant HMM for said input speech, and estimating the value of the SN ratio corresponding to that one of said incomplete noise-resistant speech HMMs which has the maximum likelihood.
Claim 20
Independent Claim
44 words
The method of any one of claims 1 to 9, 12 or 13, which further comprises a step of recording noise in an actual environment which affects the speech to be recognized, and preparing therefrom said noise HMMs for use in said first step.
Claim 21
Independent Claim
36 words
The method of any one of claims 1 to 3, 5, 6, 8 or 9, wherein said speech HMMs are represented in said cepstral domain and said noise HMMs are represented in a logarithm spectral domain.
Claim 22
Dependent Claim
329 words
A speech recognizer which uses noise-resistant speech HMMs composed of speech HMMs unaffected by noise or multiplicative distortion and noise HMMs, comprising: a speech HMM storage part having stored therein a set of said speech HMMs; a noise HMM storage part having stored therein said noise HMMS; a noise-resistant speech HMM storage part for storing said set of noise-resistant speech HMMs; a multiplicative distortion or SN ratio storage part having stored therein multiplicative distortions or SN ratios which are unknown variables; noise input means for recording noise of the same environment as that of the speech to be recognized; noise HMM preparing means for preparing said noise HMMs from said noise recorded by said noise input means and for storing said noise HMMs in said noise storage part; incomplete noise-resistant speech HMM preparing means for preparing incomplete noise-resistant speech HMMs each containing said multiplicative distortion or SN ratio as an unknown variable from said noise HMMs stored in said noise HMM storage part, said speech HMMs stored in said speech HMM storage part and said multiplicative distortions or SN ratios stored in said multiplicative distortion or SN ratio storage part; speech input means for inputting therethrough said speech to be recognized; multiplicative distortion or SN ratio estimating means for estimating said multiplicative distortion or SN ratio corresponding to that one of said incomplete noise-resistant speech HMMs which has the maximum likelihood for said speech inputted via said speech input means; noise-resistant speech HMM composing means for composing said set of noise-resistant speech HMMs by substituting said estimated multiplicative distortion or SN ratio into said incomplete noise-resistant speech HMMs and for storing said set of noise-resistant speech HMMs in said noise-resistant speech HMM storage part; and speech recognizing means for calculating the similarity between said speech to be recognized, inputted via said speech input means, and each of said noise-resistant speech HMMs in said noise-resistant speech HMM storage part and for outputting the result of recognition based on said calculation.
Claim 23
Dependent Claim
329 words
A speech recognizer which uses noise-resistant speech HMMs composed of speech HMMs unaffected by noise or multiplicative distortion and noise HMMs, comprising: a speech HMM storage part having stored therein a set of said speech HMMs; a noise HMM storage part having stored therein said noise HMMS; a noise-resistant speech HMM storage part for storing said set of noise-resistant speech HMMs; a multiplicative distortion or SN ratio storage part having stored therein multiplicative distortions or SN ratios which are unknown variables; noise input means for recording noise of the same environment as that of the speech to be recognized; noise HMM preparing means for preparing said noise HMMs from said noise recorded by said noise input means and for storing said noise HMMs in said noise storage part; incomplete noise-resistant speech HMM preparing means for preparing incomplete noise-resistant speech HMMs each containing said multiplicative distortion or SN ratio as an unknown variable from said noise HMMs stored in said noise HMM storage part, said speech HMMs stored in said speech HMM storage part and said multiplicative distortions or SN ratios stored in said multiplicative distortion or SN ratio storage part; speech input means for inputting therethrough said speech to be recognized; multiplicative distortion or SN ratio estimating means for estimating said multiplicative distortion or SN ratio corresponding to that one of said incomplete noise-resistant speech HMMs which has the maximum likelihood for said speech inputted via said speech input means; noise-resistant speech HMM composing means for composing said set of noise-resistant speech HMMs by substituting said estimated multiplicative distortion or SN ratio into said incomplete noise-resistant speech HMMs and for storing said set of noise-resistant speech HMMs in said noise-resistant speech HMM storage part; and speech recognizing means for calculating the similarity between said speech to be recognized, inputted via said speech input means, and each of said noise-resistant speech HMMs in said noise-resistant speech HMM storage part and for outputting the result of recognition based on said calculation.
References Cited (Backward Citations)
US4933973A
Cited
US5590242A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2001-02-23 | 2001-08-25 | 2002-02-24 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2005-02-23 | 2005-08-25 | 2006-02-24 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2009-02-23 | 2009-08-25 | 2010-02-24 | $7700.00 | Window Closed / Lapsed |