Method of speech recognition resistant to convolutive distortion and additive distortion

• US 20030115055A1
• Filed: 09/20/2002
• Published: 06/19/2003
• Est. Priority Date: 12/12/2001
• Status: Active Grant

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Gong, Yifan
• Granted Patent: US 7,165,028 B2
• Time in Patent Office: Days
• US Class Current: 704/233
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence

Abstract

A speech recognizer operating in both ambient noise (additive distortion) and microphone changes (convolutive distortion) is provided. For each utterance to be recognized the recognizer system adapts HMM mean vectors with noise estimates calculated from pre-utterance pause and a channel estimate calculated using an Estimation Maximization algorithm from previous utterances.

Claims

1. A speech recognizer comprising:

a microphone for receiving speech utterances and pauses;

models for recognizing speech;

a model adjustor responsive to said models for providing adjusted models;

said adjusted models having their parameters simultaneously adjusted to handle distortion introduced by convolutive and additive components; and

a recognizer coupled to said microphone and responsive to adjusted models for recognizing speech.

2. The recognizer of claim 1 wherein said model adjustor identifies and measures channel distortion and additive distortion from incoming signals received at said microphone during speech and non-speech segments.

3. The recognizer of claim 2 wherein said additive distortion is determined by calculating the average of log-domain pre-utterance input frames.

4. The recognizer of claim 2 wherein the convolutive distortion is calculated iteratively and the new convolutive distortion is calculated by subtracting a log-domain bias from the current one.

5. The recognizer of claim 4 wherein said bias is a normalized difference between compensated mean vectors that are weighted by their closeness to the input frames.

6. The recognizer of claim 5 wherein said recognizer compensates for the distortions by modifying mean vectors by compensating static MFCC, compensating first order dynamic MFCC and compensating for second order MFCC.

7. The recognizer of claim 1 wherein said model adjustor includes compensating for the distortions by modifying mean vectors by compensating static MFCC, compensating first order dynamic MFCC and compensating for second order MFCC.

8. The recognizer of claim 1 wherein said model adjustor includes recognizing speech utterances one by one and performing adjustment of internal parameters at the same time without additional data collection.

9. The recognizer of claim 1 wherein said model adjustor identifies and measures the convolutive and additive components from incoming signals during speech and non-speech segments.

10. The recognizer of claim 9 wherein said model ajustor uses measured convolutive and additive components from past utterances in determining a current utterance.

11. A method of speech recognition comprising the steps of:

receiving input speech and non-speech at a microphone, providing initial HMM models;

recognizing speech utterances one by one and performing adjustment of internal parameters at the same time without additional data collection wherein said recognizing step includes modifying mean vectors of said HMM models with current estimates compensating for static MFCC, first order dynamic MFCC and second order MFCC.

12. The method of claim 11 wherein said recognizing step includes the step of calculating the channel or microphone component iteratively and in each step the new channel component is calculated by subtracting a log-domain bias from the current one.

13. The method of claim 12 wherein said bias is a normalized difference between compensated mean vectors that are weighted by their closeness to the input frames.

14. A speech recognizer operating in both ambient noise (additive distortion) and channel changes (convolutive distortion) comprising:

HMM models and

a model adjustor coupled to a recognizer responsive to said HMM models for each utterance to be recognized adapting HMM mean vectors with noise estimates calculated from pre-utterance pause and a channel estimate calculated using an Estimation Maximization algorithm from previous utterances.

15. A method of speech recognition comprising the steps of:

receiving input speech and non-speech at a microphone, providing initial HMM models;

recognizing speech utterances one by one and performing adjustment of internal parameters for convolutive distortion and additive distortion at the same time without additional data collection.

16. The method of claim 15 wherein said recognizing step includes modifying mean vectors of said HMM models with current estimates compensating for static MFCC, first order dynamic MFCC and second order MFCC.

17. The method of claim 15 including the step of carrying the quantities allowing the determination of channel estimate obtained during determination of past utterances to a current utterance.