Speaker adaptation using discriminative linear regression on time-varying mean parameters in trended HMM

• US 6,112,175 A
• Filed: 03/02/1998
• Issued: 08/29/2000
• Est. Priority Date: 03/02/1998
• Status: Expired due to Fees

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Chengalvarayan, Rathinavelu
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US09/032,902
• Time in Patent Office: 911 Days
• Field of Search: 704/255, 704/252, 704/256
• US Class Current: 704/256.5
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A method and apparatus using a combined MLLR and MCE approach to estimating the time-varying polynomial Gaussian mean functions in the trended HMM has advantageous results. This integrated approach is referred to as the minimum classification error linear regression (MCELR), which has been developed and implemented in speaker adaptation experiments using a large body of utterances from different types of speakers. Experimental results show that the adaptation of linear regression on time-varying mean parameters is always better when fewer than three adaptation tokens are used.

Claims

1. A model-space transformation method for speech recognition comprising the steps of:

applying a linear transformation to a trended HMM; and

estimating mean parameters of said trended HMM with a minimum classification error (MCE) training;

said linear transform applying step adapting a set of speaker independent models to a

specific speaker by applying a set of linear transformations to respective Gaussian mean vectors thereof.

2. The method of claim 1 wherein:

said estimating step includes optimally estimating linear transformations to a set of time-varying mean parameters in the trended HMM.

3. The method of claim 1 further comprising the steps of:

said estimating step includes MCE training that optimally estimates a block-diagonal transformation to a set of time-varying mean parameters in the trended HMM;

wherein diagonal blocks of said block-diagonal transformation separately transform static and dynamic features and off diagonal blocks have no effect.

4. The method of claim 3 wherein said block-diagonal transformation leads to fewer parameters and more robust estimations with limited amounts of adaption data.

5. The method of claim 4, wherein said limited amounts of adaption data is at least one token of adaption data.

6. The method of claim 4, wherein said limited amounts of adaption data is at least one token and less than four tokens of adaption data.

7. A method of recognizing speech, comprising the steps of:

applying global linear regression based speaker adaption using MCE-based discriminative training in a trended HMM;

optimally estimating time varying polynomial Gaussian mean functions in the trended HMM; and

sequentially performing modified viterbi score calculations for each alphabet classification task.