Unsupervised HMM adaptation based on speech-silence discrimination

• US 6,076,057 A
• Filed: 05/21/1997
• Issued: 06/13/2000
• Est. Priority Date: 05/21/1997
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Narayanan, Shrikanth Sambasivan, Potamianos, Alexandros, Zeljkovic, Ilija
• Primary Examiner(s): Voeltz, Emanuel Todd
• Assistant Examiner(s): SOFOCLEOUS, MICHAEL D
• Application Number:
  US08/861,413
• Time in Patent Office: 1,119 Days
• Field of Search: 704/210, 704/214, 704/244, 704/256, 704/231, 704/233
• US Class Current: 704/256.2
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

An unsupervised, discriminative, sentence level, HMM adaptation based on speech-silence classification is presented. Silence and speech regions are determined either using a speech end-pointer or the segmentation obtained from the recognizer in a first pass. The discriminative training procedure using a GPD or any other discriminative training algorithm, employed in conjunction with the HMM-based recognizer, is then used to increase the discrimination between silence and speech.

Claims

1. A method for discriminating between speech and background regions, comprising:

segmenting an input utterance into speech and background regions without knowledge of the lexical content of the input utterance to create a segmented input string;

introducing insertion errors into the background regions that are error prone to generate error laden background strings;

statistically modeling the segmented input string and the error laden background strings using a discriminative training algorithm to generate a model with adapted parameters;

decoding the input utterance using the model with the adapted parameters; and

outputting a recognized string based on the decoding step.

2. The method as recited in claim 1, wherein the model with adapted parameters is generated using Hidden Markov Models.

3. The method as recited in claim 1, wherein the segmenting step uses Viterbi decoding.

4. The method as recited in claim 1, wherein the discriminative training algorithm is a minimum string-error training algorithm using N competing string models.

5. The method as recited in claim 1, wherein the decoding step uses Viterbi decoding.

6. The method as recited in claim 1, wherein the statistically modeling step uses a Generalized Probabilistic Descent algorithm.

7. A system for decoding of speech information comprising:

means for segmenting an input utterance into speech and background regions without knowledge of the lexical content of the input utterance to create a segmented input string;

means for introducing insertion errors into the background regions that are error prone to generate error laden background strings;

means for statistically modeling the segmented input string and the error laden background strings using a discriminative training algorithm to generate a model with adapted parameters;

means for decoding the input utterance using the model with the adapted parameters; and

means for outputting a recognized string based on the decoded input utterance.

8. The system as recited in claim 7, wherein the means for statistically modeling generates the model with adapted parameters using Hidden Markov Models.

9. The system as recited in claim 7, wherein the means for segmenting the input utterances into the segments uses Viterbi decoding.

10. The system as recited in claim 7, wherein the discriminative training algorithm includes a minimum string-error training algorithm using N competing string models.

11. The system as recited in claim 7, wherein means for decoding the input utterance uses Viterbi decoding.

12. The system as recited in claim 7, wherein the means for statistically modeling uses a Generalized Probabilistic Descent algorithm.