System and method for lossy compression of voice recognition models

• US 6,681,207 B2
• Filed: 01/12/2001
• Issued: 01/20/2004
• Est. Priority Date: 01/12/2001
• Status: Expired due to Term

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Garudadri, Harinath
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US09/760,076
• Publication Number:
  US 20020133345A1
• Time in Patent Office: 1,103 Days
• Field of Search: 704/230, 704/241, 704/251, 704/256, 704/246
• US Class Current: 704/256
• CPC Class Codes: G10L 15/06 Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice G10L15/14 takes precedence

Abstract

A method and system that improves voice recognition by improving storage of voice recognition (VR) templates. The improved storage means that more VR models can be stored in memory. The more VR models that are stored in memory, the more robust the VR system and therefore the more accurate the VR system. Lossy compression techniques are used to compress VR models. In one embodiment, A-law compression and A-law expansion are used to compress and expand VR models. In another embodiment, Mu-law compression and Mu-law expansion are used to compress and expand VR models. VR models are compressed during a training process and they are expanded during voice recognition.

Claims

1. A method of voice recognition (VR), comprising:

recording a plurality of utterances;

extracting features of the plurality of utterances to generate extracted features of the plurality of utterances;

creating a plurality of VR models from the extracted features of the plurality of utterances; and

lossy-compressing the plurality of VR models using A-law compression to quantize information bits of the plurality of VR models.

3. The method of claim 1 or 2, wherein the plurality of VR models are Hidden Markov Models (HMMs).

4. The method of claim 1 or 2, wherein the plurality of VR models are Dynamic Time Warping (DTW) models.

5. The method of claim 1, further comprising expanding an A-law compressed VR models from the plurality of A-law compressed VR models to generate an expanded VR model.

6. The method of claim 5, further comprising extracting features of a test utterance.

7. The method of claim 6, further comprising matching the extracted features of the utterance to an expanded VR model of the plurality of expanded VR models to generate a match.

8. The method of claim 7, further comprising generating a hypothesis for the match.

2. A method of voice recognition (VR), comprising:

recording a plurality of utterances;

extracting features of the plurality of utterances to generate extracted features of the plurality of utterances;

creating a plurality of VR models from the extracted features of the plurality of utterances; and

lossy-compressing the plurality of VR models using mu-law compression to quantize information bits of the plurality of VR models.

9. The method of claim 2, further comprising expanding a mu-law compressed VR model from the plurality of mu-law compressed models to generate an expanded VR model.

10. The method of claim 9, further comprising extracting features of a test utterance.

11. The method of claim 10, further comprising matching the extracted features of the test utterance to an expanded VR models to generate a match.

12. The method of claim 11, further comprising generating ahypothesis for the match.

13. A voice recognition (VR) system comprising a training module configured to extract features of a plurality of utterances to generate extracted features of the utterances, create a plurality of VR models from the extracted features of the utterances, and lossy-compress the plurality of VR models using A-law compression to quantize information bits of the plurality of VR models.

14. The VR system of claim 13, further comprising:

a feature extraction module configured to extract features of a test utterance to generate extracted features of a test utterance;

an expansion module configured to expand a lossy-compressed VR model from the plurality of lossy-compressed VR models to generate an expanded VR model; and

a pattern-matching module that matches the extracted features of the test utterance to the expanded VR model to generate a recognition hypothesis.

15. A voice recognition (VR) system, comprising:

a plurality of lossy-compressed VR models, wherein A-law compression is used to quantize the information bits of the VR models;

a feature extraction module configured to extract features of a test utterance to generate extracted features of a test utterance;

an expansion module configured to expand a lossy-compressed VR model from

the plurality of lossy-compressed VR models to generate an expanded VR model; and

a pattern-matching module that matches the extracted features of the test utterance to the expanded VR model to generate a recognition hypothesis.

16. A voice recognition (VR) training system comprising:

a feature extraction module configured to extract features of a plurality of utterances and generate a plurality of VR models for the extracted features of the plurality of utterances; and

a compression module configured to lossy-compress the plurality of VR models using mu-law compression to quantize information bits of the plurality of VR models.