High-accuracy, low-distortion time-frequency analysis of signals using rotated-window spectrograms

• US 5,845,241 A
• Filed: 09/04/1996
• Issued: 12/01/1998
• Est. Priority Date: 09/04/1996
• Status: Expired due to Term

• Current Assignee: Hughes Electronics Corporation (AT&T, Inc.)
• Original Assignee: Hughes Electronics Corporation (AT&T, Inc.)
• Inventors: Owechko, Yuri
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/707,540
• Time in Patent Office: 818 Days
• Field of Search: 704/215, 704/200, 704/205, 704/268, 704/276, 704/251, 704/203, 704/231, 704/229, 704/230
• US Class Current: 704/203
• CPC Class Codes: G06F 17/141 Discrete Fourier transforms
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band
• CPC Class Codes: G10L 25/48 specially adapted for particular use

Abstract

A speech processing and analysis apparatus and method for generating a time-frequency distribution of a speech signal combines a set of spectrograms with varying window lengths and orientations to provide a parameter-less time-frequency distribution having good joint time and frequency resolution at all angular orientations. The analysis window of a spectrogram is rotated relative to the frequency components of the signal by preprocessing using a Fractional Fourier Transform to form rotated window spectrograms. In particular, to form the rotated window spectrogram, the signal is initially pre-processed using a Fractional Fourier Transform of angle α, the spectrogram time-frequency distribution of the pre-processed signal is then computed using analysis window h(t) and then rotated by angle -α. The geometric mean of a set of rotated window spectrograms, which are indexed by both the analysis window length and the angular orientation of the window relative to the signal'"'"'s time-frequency features, is then computed to form a combination of rotated window spectrograms.

Claims

1. A speech recognition apparatus, comprising:

a spectral shaping source for generating a plurality of digital signal samples representative of an input speech signal;

a signal processor coupled to said source, comprising;

means for transforming said plurality of said signal samples to pre-processed signals representative of the frequency domain at various angular orientations;

means for generating initial time-frequency distributions of said pre-processed signals using analysis windows; and

means for rotating said time-frequency distributions back by said various angular orientations for generating a plurality of rotated window spectrograms, and

signal modeling apparatus for comparing the plurality of rotated window spectrograms against each of a plurality of word models and identifying the closest match.

2. The apparatus claimed in claim 1, wherein said processor means, further comprises:

means for combining said plurality of rotated window spectrograms to form a parameter-less combined rotated window spectrogram.

3. The apparatus claimed in claim 2, wherein said means for transforming said plurality of said signal samples to pre-processed signals representative of the frequency domain at various angular orientations, further comprises:

Fractional Fourier transform means for generating pre-processed signals representative of the frequency domain at various angular orientations.

4. The apparatus claimed in claim 3, wherein said Fractional Fourier transform means for generating pre-processed signals representative of the frequency domain at various angular orientations, further comprises:

means for pre-processing using the Fractional Fourier transform of an angle equal to the angular orientation of the window relative to said signal'"'"'s time-frequency characteristics.

5. The apparatus claimed in claim 4, wherein said means for generating initial time-frequency distributions of said pre-processed signals using analysis windows, further comprises:

spectrogram means for generating time-frequency distributions of said pre-processed signals.

6. The apparatus claimed in claim 5, wherein said means for combining said plurality of rotated window spectrograms to form a parameter-less combined rotated window spectrogram, further comprises:

means for computing a geometric mean of said plurality of rotated window spectrograms.

7. The apparatus claimed in claim 6, wherein said means for computing a geometric mean of said plurality of rotated window spectrograms, further comprises:

means for indexing the analysis window length and angular orientation of said plurality of rotated window spectrograms relative to time and frequency.

8. The apparatus claimed in claim 7, wherein said Fractional Fourier Transform means rotates the time-frequency components of said signal samples until they match 0 and π

/2.

9. The apparatus claimed in claim 8, wherein said source for generating a plurality of digital signal samples representative of said signal, further comprises:

convertor means for digitally converting said signal.

10. The apparatus claimed in claim 9, wherein said apparatus is utililized for speech processing.

11. The apparatus claimed in claim 10, further comprising:

signal modeling means for generating word models in response to said combined rotated window spectrograms.

12. The apparatus claimed in claim 11, wherein said speech processing includes speech recognition.

13. The apparatus claimed in claim 11, wherein said speech processing includes includes speaker verification.

14. The apparatus claimed in claim 9, wherein said signal is a diagnostic signal.

15. The apparatus claimed in claim 9, wherein said signal is a sonar signal.

16. The apparatus claimed in claim 9, wherein said signal is a radar signal.

17. A speech recognition method, comprising the steps of:

capturing a plurality of digital signal samples representative of an input speech signal;

generating a plurality of rotated window spectrograms from said signal samples, comprising the steps of;

transforming said plurality of said signal samples to pre-processed signals representative of the frequency domain at various angular orientations using Fractional Fourier Transform means;

generating initial time-frequency distributions of said pre-processed signals using analysis windows; and

rotating said time-frequency distributions back by said various angular orientations, and

generating word models in response to said plurality of rotated window spectrograms.

18. The method claimed in claim 17, wherein said step of generating a plurality of rotated window spectrograms further comprises the step of:

combining said plurality of rotated window spectrograms to form a parameter-less combined rotated window spectrogram.

19. The method claimed in claim 18, wherein said step of transforming comprises:

pre-processing using the Fractional Fourier Transform of an angle equal to the angular orientation of the window relative to the time-frequency characteristics of said signal samples.

20. The method claimed in claim 19, wherein said step of generating initial time-frequency distributions of said pre-processed signals using analysis windows, further comprises the step of:

generating time-frequency distributions of said pre-processed signals using spectrogram means.

21. The method claimed in claim 20, wherein said step of combining said plurality of rotated window spectrograms to form a parameter-less combined rotated window spectrogram, further comprises the step of:

computing a geometric mean of said plurality of rotated window spectrograms.

22. The method claimed in claim 21, wherein said step of computing a geometric mean of said plurality of rotated window spectrograms, further comprises the step of:

indexing the analysis window length and angular orientation of said plurality of rotated window spectrograms relative to time and frequency.

23. The method claimed in claim 22, wherein said step of transforming, further comprises the step of:

rotating the time-frequency components of said signal samples until they match 0 and π

/2 angles.

24. The method claimed in claim 23, wherein said step of capturing a plurality of digital signal samples representative of said input speech signal, further comprises the step of:

digitally converting said input speech signal.

25. The method claimed in claim 24, wherein said signal is a diagnostic signal.

26. The method claimed in claim 24, wherein said signal is a sonar signal.

27. The method claimed in claim 24, wherein said signal is a radar signal.

28. The method claimed in claim 19, wherein said step of generating word models comprises generating word models in response to said combined rotated window spectrograms.

29. The method claimed in claim 28, further comprising the step of comparing said word models generated in response to said combined rotated window spectrograms against each of a plurality of word models to identify the closest match.

30. The method claimed in claim 29, further including a step of speaker verification using the word model identified as the closest match.