Speech pattern matching in non-white noise

• US 5,487,129 A
• Filed: 06/13/1994
• Issued: 01/23/1996
• Est. Priority Date: 08/01/1991
• Status: Expired due to Term

• Current Assignee: DSP Group Incorporated (Synaptics Incorporated)
• Original Assignee: DSP Group Incorporated (Synaptics Incorporated)
• Inventors: Haimi-Cohen, Raziel, Shallom, Ilan D., Paiss, Omri, Flomen, Felix
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Doerrler, Michelle
• Application Number:
  US08/190,087
• Time in Patent Office: 589 Days
• Field of Search: 381/45, 395/2.31, 395/2.44, 395/2.35-2.37, 395/2.3, 395/2.4
• US Class Current: 704/233
• 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 system for matching an input signal, comprising non-white noise and a patterned signal corrupted by said non-white noise, to a plurality of reference signals, the system including an estimator estimating noise features of said non-white noise and producing from the noise features at least one noise whitening filter; a filter generally simultaneously filtering the input signal and the plurality of reference signals using the noise whitening filter and producing a filtered input signal, having a white noise component, and a plurality of filtered reference signals; and a pattern matcher generally robust to white noise for matching the filtered input signal to one of the filtered reference signals.

Claims

1. A system for matching an input signal, comprising non-white noise and a patterned signal corrupted by said non-white noise, to a plurality of reference signals, said system comprising:

an estimator estimating noise features of said non-white noise and producing from said noise features at least one noise whitening filter;

a filter filtering said input signal and said plurality of reference signals using said at least one noise whitening filter and producing a filtered input signal, having a white noise component, and a plurality of filtered reference signals; and

a pattern matcher for matching said filtered input signal to one of said filtered reference signals.

2. System according to claim 1 and wherein said at least one noise whitening filter is two noise whitening filters respectively for filtering the input signal and the reference signals.

3. System according to claim 1 including an extractor extracting features of said input signal and said reference signals and wherein said filter operates in a feature domain.

4. System according to claim 3 and wherein said feature domains are selected from the group of data samples, Linear Prediction Coefficients, cepstral coefficients, power spectrum samples, and filter bank energies.

5. System according to claim 4 and wherein said estimator estimates a filter in accordance with the selected feature domain.

6. System according to claim 1 and wherein a feature domain of the input signal is different than a feature domain of the reference signals.

7. System according to claim 6 and wherein said feature domains are selected from the group of data samples, Linear Prediction Coefficients, cepstral coefficients, power spectrum samples, and filter bank energies.

8. System according to claim 1 and wherein the pattern matcher performs a pattern matching technique selected from the group of Dynamic Time Warping DTW, Hidden Markov Models HMM, or Dynamic Time Warping-Vector Quantization DTW-VQ.

9. System according to claim 1 and wherein the input signal is a speech signal.

10. System according to claim 1 and wherein said non-white noise is the noise from the environment of a movable vehicle.

11. System according to claim 1 and wherein said non-white noise is the noise from the environment of a moving airplane cockpit.

12. System according to claim 1 and wherein said non-white noise is the noise from the environment of a vibrating machine.

14. Speech recognition system for recognizing a word found in a speech signal corrupted by non-white noise, said system comprising:

a vector quantization system according to claim 10 producing vector quantized speech; and

a word matcher receiving a plurality of reference sequences of symbols relating to said reference signals and a test sequence of symbols relating to said speech signal for matching said test sequence to said reference sequences thereby to recognize said word in said speech signal.

15. A speech recognition system according to claim 14 wherein said word matcher performs Dynamic Time Warping (DTW) on said vector quantized speech.

16. A speaker verification system using the system of claim 15 wherein said reference signals comprises at least one word spoken by one speaker.

17. A speech recognition system according to claim 14 wherein said word matcher performs Hidden Markov Modeling (HMM).

18. A speaker verification system using the system of claim 17 wherein said reference signals comprises at least one word spoken by one speaker.

19. A speaker recognition system using the system of claim 14 wherein said reference signals comprise one word spoken by a plurality of different speakers.

20. A speaker verification system using the system of claim 14 wherein said reference signals comprises at least one word spoken by one speaker.

13. Speech recognition system for recognizing words found in a speech signal corrupted by non-white noise, said system comprising:

an estimator estimating noise features of said non-white noise and for producing from said noise features at least one noise whitening filter;

a filter filtering said speech signal and a plurality of reference signals of selected spoken words, said filter using said at least one noise whitening filter and producing a filtered speech signal and a plurality of filtered reference signals; and

a pattern matcher for matching said filtered speech signal to one of said filtered reference signals thereby recognizing the word in said speech signal.

21. A speaker recognition system using the system of claim 13 wherein said reference signals comprise one word spoken by a plurality of different speakers.

22. A speaker verification system using the system of claim 13 wherein said reference signals comprises at least one word spoken by one speaker.

23. A Vector Quantization (VQ) system for vector quantizing a speech signal corrupted by non-white noise into a sequence of symbols, said system comprising:

an estimator estimating noise features of said non-white noise and for producing from said noise features at least one noise whitening filter;

a filter filtering segments of said speech signal and a plurality of numbered reference signals of selected speech segments, said filter using said at least one noise whitening filter and producing filtered speech segments and a plurality of filtered reference segments; and

a pattern matcher for matching each of said filtered speech segments to one of said filtered reference segments and for providing as an output a symbol which is an index of the matched reference segment.

24. A speaker recognition system using the system of claim 23 wherein said reference signals comprise one word spoken by a plurality of different speakers.

25. A speaker verification system using the system of claim 23 wherein said reference signals comprises at least one word spoken by one speaker.

26. A method for matching an input signal, comprising non-white noise and a patterned signal corrupted by said non-white noise, to a plurality of reference signals, said method comprising the steps of:

estimating noise features of said non-white noise and producing from said noise features at least one noise whitening filter;

filtering said input signal and said plurality of reference signals using said at least one noise whitening filter and producing a filtered input signal, having a white noise component, and a plurality of filtered reference signals; and

matching said filtered input signal to one of said filtered reference signals.

27. A method according to claim 26 and wherein said at least one noise whitening filter is two noise whitening filters respectively for filtering the input signal and the reference signals.

28. A method according to claim 26 including the steps of extracting features of said input signal and said reference signals and wherein said step of filtering occurs in a feature domain.

29. A method according to claim 28 and wherein said feature domains are selected from the group of data samples, Linear Prediction Coefficients, cepstral coefficients, power spectrum samples, and filter bank energies.

30. A method according to claim 28 and wherein said step of estimating estimates a filter in accordance with the selected feature domain.

31. A method according to claim 26 and wherein a feature domain of the input signal is different than a feature domain of the reference signals.

32. A method according to claim 26 and wherein said step of matching performs a pattern matching technique selected from the group of Dynamic Time Warping DTW, Hidden Markov Models HMM, or Dynamic Time Warping-Vector Quantization DTW-VQ.

33. A method according to claim 26 and wherein the input signal is a speech signal.