Linear trajectory models incorporating preprocessing parameters for speech recognition

• US 6,076,058 A
• Filed: 03/02/1998
• Issued: 06/13/2000
• 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:
  US09/032,900
• Time in Patent Office: 834 Days
• Field of Search: 704/243, 704/256, 704/202, 704/201
• US Class Current: 704/256.8
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum
• CPC Class Codes: G10L 25/27 characterised by the analysis technique

Abstract

The proposed model aims at finding an optimal linear transformation on the Mel-warped DFT features according to the minimum classification error (MCE) criterion. This linear transformation, along with the (NSHMM) parameters, are automatically trained using the gradient descent method. An advantageous error rate reduction can be realized on a standard 39-class TIMIT phone classification task in comparison with the MCE-trained NSHMM using conventional preprocessing techniques.

Claims

1. A method for speech recognition comprising the steps of:

digitizing and framing a speech utterance token;

transmitting said digitized and framed speech utterance token to Mel-filter banks;

Mel-filtering the digitized and framed utterance token to produce log energy vectors for the number of classes C;

transmitting a sequence of log energy vectors according to the frames to compute feature transformation operation;

computing feature transformations for each class i of the utterance token and transmitting the result to next operation;

computing static and dynamic features therefrom and transmitting the result to next operation;

calculating a respective log likelihood for each of the utterance tokens Pi and transmitting the result to next operation;

testing the token Pi to see if it is less than S and if it is, then the method branches to the next testing operation and if Pi is equal to S then setting index j equal to class index i and proceeding to the next testing operation;

testing to see if index i is less than the number of classes C, if yes then iterating the index i by one and proceeding back to the computing feature transformations for each class i step and repeating this iteration and return until the expression i<

C is false, which means all classes i have been processed;

if index i is not less than the number of classes C, then classification of this utterance token is finished and a j-th class is recognized for this given token;

testing to see if this is the last utterance token to be processed and if it is the last then proceeding to done, otherwise returning to the digitizing and framing step to begin processing a subsequent utterance token.

2. An apparatus for speech recognition comprising:

means for determining a minimum classification error criterion; and

means for determining an optimum transformed nonstationary state hidden Markov model according to the minimum classification error (MCE) criterion;

wherein said means for determining an optimum transformed nonstationary state hidden Markov model includes a linear transformation based on Mel-warned DFT features.

3. The apparatus of claim 2, further comprising:

means for automatically training said optimum transformed nonstationary state hidden Markov model.

4. The apparatus of claim 3, wherein said means for automatically training uses gradient decent techniques in said automatic training.