Speaker-independent model generation apparatus and speech recognition apparatus each equipped with means for splitting state having maximum increase in likelihood

• US 5,839,105 A
• Filed: 11/29/1996
• Issued: 11/17/1998
• Est. Priority Date: 11/30/1995
• Status: Expired due to Term

• Current Assignee: DENSO Corporation
• Original Assignee: ATR Interpreting Telephony Research Laboratories
• Inventors: Ostendorf, Mari, Singer, Harald
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): ZINTEL, HAROLD ALBERT
• Application Number:
  US08/758,378
• Time in Patent Office: 718 Days
• Field of Search: 704/255, 704/256, 704/257, 704/244, 704/245, 704/231
• US Class Current: 704/256
• CPC Class Codes: G10L 15/144 Training of HMMs
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

There is provided a speaker-independent model generation apparatus and a speech recognition apparatus which require a processing unit to have less memory capacity and which allow its computation time to be reduced, as compared with a conventional counterpart. A single Gaussian HMM is generated with a Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers. A state having a maximum increase in likelihood as a result of splitting one state in contextual or temporal domains is searched. Then, the state having a maximum increase in likelihood is split in a contextual or temporal domain corresponding to the maximum increase in likelihood. Thereafter, a single Gaussian HMM is generated with the Baum-Welch training algorithm, and these steps are iterated until the states within the single Gaussian HMM can no longer be split or until a predetermined number of splits is reached. Thus, a speaker-independent HMM is generated. Also, speech is recognized with reference to the generated speaker-independent HMM.

Claims

1. A speaker-independent model generation apparatus comprising:

model generation means for generating a hidden Markov model of a single Gaussian distribution using a Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers, and thereafter for generating a speaker-independent hidden Markov model by iterations of splitting a state having a maximum increase in likelihood upon splitting one state in contextual or temporal domains on the hidden Markov model of the single Gaussian distribution.

2. The speaker-independent model generation apparatus as claimed in claim 1, wherein said model generation means comprises:

initial model generation means for generating an initial hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers;

search means for searching a state having a maximum increase in likelihood upon splitting one state in contextual or temporal domains on the initial hidden Markov model of the single Gaussian distribution generated by said initial model generation means;

generation means for splitting the state having the maximum increase in likelihood searched by said search means, in a contextual or temporal domain corresponding to the maximum increase in likelihood and thereafter for generating a hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm; and

control means for generating a speaker-independent hidden Markov model by iterating a process of said search means and a process of said generation means until at least one of the following conditions is satisfied;

(a) the states within the hidden Markov model of the single Gaussian distribution can no longer be split; and

(b) a number of states within the hidden Markov model of the single Gaussian distribution reaches a predetermined number of splits.

3. The speaker-independent model generation apparatus as claimed in claim 2, wherein the states searched by the search means are limited to two new states split by said generation means in the preceding process.

4. The speaker-independent model generation apparatus as claimed in claim 2,

wherein the states searched by the search means are limited to two new states split by said generation means in the preceding step and a state which is away from the two new states by a distance of one.

5. A speech recognition apparatus comprising:

model generation means for generating a hidden Markov model of a single Gaussian distribution using a Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers, and thereafter for generating a speaker-independent hidden Markov model by iterations of splitting a state having a maximum increase in likelihood upon splitting one state in contextual or temporal domains on the hidden Markov model of the single Gaussian distribution; and

speech recognition means for, in response to an input speech signal of a spoken speech, recognizing the spoken speech with reference to the speaker-independent hidden Markov model generated by said model generation means.

6. A speech recognition apparatus comprising:

initial model generation means for generating a hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers;

search means for searching a state having a maximum increase in likelihood upon splitting one state in contextual or temporal domains on the hidden Markov model of the single Gaussian distribution generated by said initial model generation means;

generation means for splitting the state having the maximum increase in likelihood searched by said search means, in a contextual or temporal domain corresponding to the maximum increase in likelihood and thereafter for generating a hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm;

control means for generating a speaker-independent hidden Markov model by iterating a process of said search means and a process of said generation means until at least one of the following conditions is satisfied;

(a) the states within the hidden Markov model of the single Gaussian distribution can no longer be split; and

(b) a number of states within the hidden Markov model of the single Gaussian distribution reaches a predetermined number of splits; and

speech recognition means for, in response to an input speech signal of a spoken speech, recognizing the spoken speech with reference to the speaker-independent hidden Markov model generated by said control means.

7. The speech recognition apparatus as claimed in claim 6,

wherein the states searched by the search means are limited to two new states split by said generation means in the preceding process.

8. The speech recognition apparatus as claimed in claim 6,

wherein the states searched by the search means are limited to two new states split by said generation means in the preceding step and a state which is away from the two new states by a distance of one.

9. A method for generating a speaker-independent model, including the following steps:

generating a hidden Markov model of a single Gaussian distribution using a Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers; and

thereafter, generating a speaker-independent hidden Markov model by iterations of splitting a state having a maximum increase in likelihood upon splitting cone state in contextual or temporal domains on the hidden Markov model of the single Gaussian distribution.

10. A method for generating a speaker-independent model, including the following steps:

generating an initial hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm based on spoken speech data from a plurality of specific speakers;

searching a state having a maximum increase in likelihood upon splitting one state in contextual or temporal domains on the generated initial hidden Markov model of the single Gaussian distribution;

splitting the searched state having the maximum increase in likelihood, in a contextual or temporal domain corresponding to the maximum increase in likelihood, and thereafter, generating a hidden Markov model of a single Gaussian distribution using the Baum-Welch training algorithm; and

generating a speaker-independent hidden Markov model by iterating said searching step and said splitting and generating step until at least one of the following conditions is satisfied;

(a) the states within the hidden Markov model of the single Gaussian distribution can no longer be split; and

(b) a number of states within the hidden Markov model of the single Gaussian distribution reaches a predetermined number of splits.