Formant frequency estimation method, apparatus, and medium in speech recognition

• US 20070192088A1
• Filed: 01/04/2007
• Published: 08/16/2007
• Est. Priority Date: 02/10/2006
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Jeong, Jae-Hoon, Jeong, So-Young, Oh, Kwang Cheol
• Granted Patent: US 7,818,169 B2
• Time in Patent Office: Days
• US Class Current: 704/209
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 25/15 the extracted parameters being formant information
• CPC Class Codes: G10L 25/90 Pitch determination of speech signals

Abstract

A formant frequency estimation method which is important information in speech recognition by accelerating a spectrum using a pitch frequency, and an apparatus using the method is provided. That is, the formant frequency estimation method includes preprocessing an input speech signal and generating a spectrum by a fast Fourier transforming the preprocessed input speech signal; smoothing the generated spectrum; accelerating the smoothed spectrum; and determining a formant frequency on the basis of the accelerated spectrum.

Claims

1. A formant frequency estimation method in speech recognition comprising:

preprocessing an input speech signal and generating a spectrum by fast Fourier transforming the preprocessed input speech signal;

smoothing the generated spectrum;

accelerating the smoothed spectrum; and

determining a formant frequency on the basis of the accelerated spectrum.

2. The method of claim 1, further comprising

estimating a pitch frequency of the preprocessed input speech signal,

wherein the smoothing of the generated spectrum is based on a moving average of the generated spectrum and smoothes the generated spectrum by using a number of tabs corresponding to the estimated pitch frequency.

3. The method of claim 1, wherein the accelerating the smoothed spectrum comprises:

calculating a first spectral difference for the smoothed spectrum;

smoothing a spectrum of the first spectral difference; and

calculating a second spectral difference for the smoothed spectrum of the first spectral difference.

4. The method of claim 1, wherein the determining a formant frequency comprises:

selecting each domain above zero in the accelerated spectrum as a formant candidate;

calculating an auto-correlation for each of the formant candidates by using a spectral difference calculated from the accelerated spectrum; and

determining the formant candidate of which the auto-correlation is higher than a predetermined threshold value as the formant frequency.

5. The method of claim 4, wherein the determining the formant candidate determines the formant frequency by using parabolic interpolation.

10. A medium comprising computer readable instructions for implementing the method of claim 1.

6. A formant frequency estimation method in speech recognition comprising:

establishing a flag state backward;

calculating an anchor parameter after preprocessing an input speech signal;

executing buffering until the anchor parameter is above a predetermined threshold value;

estimating a backward formant frequency after the anchor parameter is above the predetermined threshold value; and

changing the flag state and establishing after estimating the backward formant frequency.

7. The method of claim 6, wherein the method further comprises estimating a forward formant frequency when the anchor parameter is above the predetermined threshold value after the flag state has changed.

8. The method of claim 7, wherein the method further comprises, after the flag state is changed, establishing the flag state backward when the anchor parameter becomes smaller than the predetermined threshold value.

9. The method of claim 6, the anchor parameter is a type of parameter for a voicing estimation.

18. A medium comprising computer readable instructions implementing the method of claim 6.

11. A formant frequency estimation apparatus in speech recognition comprising:

a preprocess unit pre-processing an input speech signal;

a fast Fourier transformation unit Fourier transforming the preprocessed input speech signal, and generating a spectrum;

a smoothing unit smoothing the generated spectrum;

an acceleration unit accelerating the smoothed spectrum; and

a formant frequency determination unit determining a formant frequency on the basis of the accelerated spectrum.

The apparatus of claim 11, wherein the formant frequency determining unit further comprises;

a pitch frequency estimation unit estimating a pitch frequency of a preprocessed input speech signal, and

wherein the smoothing unit is based on a moving average of the generated spectrum and smoothes the generated spectrum by using the number of tabs corresponding to the estimated pitch frequency.

14. The apparatus of claim 11, the formant frequency determination unit selects each domain in which it'"'"'s domain is higher than a zero in the accelerated spectrum, as a formant candidate, calculates an auto-correlation in the each formant candidate by using the calculated spectral difference which is calculated by a spectrum difference in the accelerated spectrum, and determines a formant candidate of which the auto-correlation is higher than a predetermined threshold value as the formant frequency.

13. The apparatus of claim 12, wherein the acceleration unit calculates a first spectral difference corresponding to the smoothed spectrum, smoothes a spectrum of the first spectral difference, and calculates a second spectral difference corresponding to the spectrum the smoothed first spectral difference.

15. A formant frequency estimation apparatus in speech recognition comprising:

a flag state establishment unit establishing a flag state backward;

an anchor parameter calculation unit calculating an anchor parameter after a preprocess of input speech signal;

a buffering unit executing buffering until the anchor parameter is above a predetermined threshold value;

a formant frequency estimation unit estimating a backward formant frequency after anchor parameter is above the predetermined threshold value; and

wherein the flag state establishment unit changes and establishes the flag state after estimating the backward formant frequency.

16. The apparatus of claim 15, wherein the formant frequency estimation apparatus estimates a forward formant frequency when the anchor parameter is above to the predetermined threshold value after the flag state has changed, and

the flag state establishment unit repeatedly establishes the flag state backward when the anchor parameter become smaller than the predetermined threshold value after the flag state has changed.

17. The apparatus of claim 15, wherein the anchor parameter is a type of parameter for a voicing estimation.

19. A formant frequency estimation method in speech recognition comprising:

calculating an anchor parameter from an input speech signal;

executing buffering until the anchor parameter is above a predetermined threshold value; and

estimating a backward formant frequency after the anchor parameter is above the predetermined threshold value.

20. The method of claim 19, further comprising establishing a flag state backward before calculating the anchor parameter.

21. The method of claim 20, further comprising changing the flag state and establishing a flag state not-backward after estimating the backward formant frequency.

22. The method of claim 21, further comprising estimating a forward formant frequency when the anchor parameter is above the predetermined threshold value after the flag state has changed.

23. The method of claim 22, further comprising:

after the flag state is changed, establishing the flag state backward when the anchor parameter becomes smaller than the predetermined threshold value.

24. A medium comprising computer readable instructions implementing the method of claim 19.