Method and system for adaptive speech recognition in a noisy environment

• US 6,418,411 B1
• Filed: 02/10/2000
• Issued: 07/09/2002
• Est. Priority Date: 03/12/1999
• Status: Active Grant

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Gong, Yifan
• Primary Examiner(s): Banks-Harold, Marsha D.
• Assistant Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US09/501,885
• Time in Patent Office: 880 Days
• Field of Search: 704/256, 704/255, 704/226
• US Class Current: 704/256.5
• CPC Class Codes: G10L 15/144 Training of HMMs
• 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

The system uses utterances recorded in low noise condition, such as a car engine off to optimally adapt speech acoustic models to transducer and speaker characteristics and uses speech pauses to adjust the adopted models to a changing background noise, such as when in a car with the engine running.

Claims

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

performing a one time channel adaptation of initial Hidden Markov Models for all successive speech utterances to provide channel adapted Hidden Markov Models;

determining background noise for each utterance in a time period adjacent speech to be recognized;

adapting channel adapted Hidden Markov Models to the determined background noise for each utterance to get channel and noise adapted target Hidden Markov Models; and

recognizing speech adjacent to the determined background noise using said channel and noise adapted target Hidden Markov Models.

2. The method of claim 1 including the step of performing one time gender adaptation of said Hidden Markov Models to provide channel and gender adapted Hidden Markov Models, said adapting step adapts channel and gender adapted Hidden Markov Models to noise to get channel, gender and noise adapted target Hidden Markov Models and said recognizing step uses channel, gender, and noise adapted target Hidden Markov Models.

3. The method of claim 1 including the step of performing one time speaker adaptation of said Hidden Markov Models to provide channel and speaker adapted Hidden Markov Models, said adapting step adapts channel and speaker adapted Hidden Markov Models to noise to get channel, speaker and noise adapted target Hidden Markov Models, said recognizing step uses said channel speaker and noise adapted Hidden Markov Models.

4. The method of claim 3 including the step of providing one time gender adaptation of said Hidden Markov Models to provide channel, gender, and speaker adapted Hidden Markov Models, said adapting step adapts channel, gender, and speaker adapted Hidden Markov Models to noise to get channel, gender, speaker and noise adapted Hidden Markov Models, said recognizing step uses said channel, gender and speaker noise adapted Hidden Markov Models.

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

performing one time channel, then gender, then speaker sequence adaptation of initial Hidden Markov Models for all successive speech utterances to provide target channel, gender, and speaker adapted Hidden Markov Models;

determining background noise for each utterance in a time period adjacent speech to be recognized;

adapting target channel, gender, and speaker adapted Hidden Markov Models to the determined background noise to get channel, gender, speaker and noise adapted target Hidden Markov Models; and

recognizing speech adjacent to the determined background noise using said channel, gender, speaker and noise adapted target Hidden Markov Models.

6. The method of claim 5 wherein the step of determining background noise includes sensing noise just prior to receiving an utterance to be recognized.

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

determining background noise in a time period adjacent speech to be recognized;

adapting Hidden Markov Models to determine background noise to get noise adapted Hidden Markov Models; and

recognizing speech adjacent to the determined background noise using said noise adapted target Hidden Markov Models, said step of determining background noise is just prior to receiving an utterance to be recognized; and

said step of determining background noise includes the step of storing background noise after a key to talk switch is operated and before any utterance is received.

8. The method of claim 7 wherein said step of determining background noise includes an alarm indicator for notifying the speaker when to speak for recognition.

9. A system for speech recognition comprising:

a background noise sensor responsive to background noise of the speaker for speech recognition;

a noise adapter responsive to said background noise for adapting Hidden Markov Models to said sensed noise;

a speech recognizer responsive to the speech of said speaker and said noise adapted Hidden Markov Models for recognizing speech of said speaker; and

said sensor includes means responsive to a key to talk switch for sensing said noise and delay indicator device for notifying the speaker to speak after keying the speech to talk switch.

10. The system of claim 9 wherein said system includes adaptation of initial Hidden Markov Models in a low noise environment prior to determining background noise at said sensor.

11. The system of claim 10 wherein adaptation of initial Hidden Markov Models includes a channel adapter.

12. The system of claim 11 wherein said adaptation of initial Hidden Markov Models includes a speaker adaptation.

13. The system of claim 12 wherein said adaptation of said Hidden Markov Models includes channel adaptation, gender adaptation and speaker adaptation.