Method for extracting feature vectors for speech recognition

• US 20060129392A1
• Filed: 12/08/2005
• Published: 06/15/2006
• Est. Priority Date: 12/13/2004
• Status: Abandoned Application

• Current Assignee: LG Electronics, Inc. (LG Corporation)
• Original Assignee: LG Electronics, Inc. (LG Corporation)
• Inventors: Kim, Chan-Woo
• Application Number:
  US11/296,293
• Publication Number:
  US 20060129392A1
• Time in Patent Office: Days
• US Class Current: 704/233
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 25/93 Discriminating between voiced and unvoiced parts of speech signals G10L25/90 takes precedence

Abstract

Disclosed is a method for speech recognition which achieves a high recognition rate. The method includes extracting a parameter from an input signal that represents a characterization of the input signal as a voiced or unvoiced sound, extracting at least one feature vector corresponding to an overall spectrum shape of a voice from an input signal, and using the extracted parameter and extracted feature vectors in a training phase and in a recognition phase to recognize speech.

Claims

1. A method for recognizing speech, comprising:

extracting a parameter from an input signal that represents a characterization of the input signal as a voiced or unvoiced sound; and

recognizing speech based upon the extracted parameter.

2. The method according to claim 1, further comprising extracting a feature vector based upon the extracted parameter.

3. The method according to claim 1, wherein the parameter is calculated using the equation:

$\mathrm{\eta <br><br>}=\frac{\max <br><br>\text{\hspace{1em}<br><br>}<br><br>r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[n\right]}{r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[0\right]},\mathrm{wherein}$ $r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}^{\left(k\right)}<br><br>\left[n\right]=\frac{1}{N_f}<br><br>{\left\{\underset{n=0}{\overset{N_f-1}{\sum <br><br>}}<br><br>{<br><br>x<br><br>\left[n\right]-x<br><br>\left[n-m\right]<br><br>}^k\right\}}^{\frac{1}{k}},$

η

represents the extracted parameter, and N_f represents the length of a frame in which it is determined whether a sound is voiced or unvoiced.

4. The method according to claim 3, wherein the value of k is one of 1, 2 and 3.

5. The method according to claim 1, wherein the extracted parameter is greater than or equal to a threshold value when the input signal comprises a voiced sound.

6. The method according to claim 1, wherein the extracted parameter is less than a threshold value when the input signal comprises an unvoiced sound.

7. The method according to claim 1, wherein recognizing speech comprises utilizing a Hidden Markov Model-based recognition method.

8. The method according to claim 1, wherein recognizing speech comprises utilizing a Dynamic Time Warping-based recognition method.

9. The method according to claim 1, wherein recognizing speech comprises utilizing a neural network-based recognition method.

10. The method according to claim 1, further comprising:

generating a bit which indicates whether the input signal comprises a voiced sound or an unvoiced sound, based upon the extracted parameter, and

recognizing the speech based upon the generated bit.

11. The method according to claim 1, further comprising adding at least one of a differential coefficient and an acceleration coefficient to the extracted parameter.

12. A method for recognizing speech, comprising:

extracting at least one feature vector corresponding to an overall spectrum shape of a speech from an input signal;

extracting a parameter from the input signal that represents a characterization of the input signal as a voiced or unvoiced sound; and

recognizing speech based upon the at least one extracted feature vector and extracted parameter.

13. The method according to claim 12, wherein the parameter is calculated within an available pitch range using the equation:

$\mathrm{\eta <br><br>}=\frac{\max <br><br>\text{\hspace{1em}<br><br>}<br><br>r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[n\right]}{r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[0\right]},$

wherein

an autocorrelation function

$r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[n\right]=\underset{n=0}{\overset{N_f}{\sum <br><br>}}<br><br>{<br><br>x<br><br>\left[n\right]-x<br><br>\left[n-m\right]<br><br>}^k,$

η

represents the extracted parameter, and N_f represents the length of a frame in which it is determined whether a sound is voiced or unvoiced.

14. The method according to claim 12, wherein the extracted parameter is greater than or equal to a threshold value when the input signal comprises a voiced sound.

15. The method according to claim 12, wherein the extracted parameter is less than a threshold value when the input signal comprises an unvoiced sound.

16. The method according to claim 12, wherein recognizing speech comprises utilizing one of a Hidden Markov Model recognition method, a Dynamic Time Warping recognition method and a neural network recognition method.

17. The method according to claim 12, further comprising:

generating a bit which indicates whether the input signal comprises a voiced sound or an unvoiced sound, based upon the extracted parameter, and

recognizing the speech based upon the generated bit.

18. The method according to claim 12, further comprising adding at least one of a differential coefficient and an acceleration coefficient to the extracted parameter.

19. A computer-readable medium which comprises a program for recognizing speech, the program comprising instructions for:

extracting a parameter from an input signal that represents a characterization of the input signal as a voiced or unvoiced sound; and

recognizing speech based upon the extracted parameter.

20. The computer-readable medium according to claim 19, wherein the program further comprises instructions for extracting a feature vector based upon the extracted parameter.

21. The computer-readable medium according to claim 19, wherein the parameter is calculated using the equation:

$\mathrm{\eta <br><br>}=\frac{\max <br><br>\text{\hspace{1em}<br><br>}<br><br>r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[n\right]}{r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}<br><br>\left[0\right]},\mathrm{wherein}$ $r_{\text{\hspace{1em}<br><br>}<br><br>\underset{\_}{x}}^{\left(k\right)}<br><br>\left[n\right]=\frac{1}{N_f}<br><br>{\left\{\underset{n=0}{\overset{N_f-1}{\sum <br><br>}}<br><br>{<br><br>x<br><br>\left[n\right]-x<br><br>\left[n-m\right]<br><br>}^k\right\}}^{\frac{1}{k}},$

η

represents the extracted parameter, and N_f represents the length of a frame in which it is determined whether a sound is voiced or unvoiced.

22. The computer-readable medium according to claim 21, wherein the value of k is one of 1, 2 and 3.

23. The computer-readable medium according to claim 19, wherein the extracted parameter is greater than or equal to a threshold value when the input signal comprises a voiced sound.

24. The computer-readable medium according to claim 19, wherein the extracted parameter is less than a threshold value when the input signal comprises an unvoiced sound.

25. The computer-readable medium according to claim 19, wherein the instructions for recognizing speech comprise instructions which utilize a Hidden Markov Model-based recognition method.

26. The computer-readable medium according to claim 19, wherein the instructions for recognizing speech comprise instructions which utilize a Dynamic Time Warping-based recognition method.

27. The computer-readable medium according to claim 19, wherein the instructions for recognizing speech comprise instructions which utilize a neural network-based recognition method.

28. The computer-readable medium according to claim 19, wherein the program comprises further instructions for:

generating a bit which indicates whether the input signal comprises a voiced sound or an unvoiced sound, based upon the extracted parameter, and

recognizing the speech based upon the generated bit.

29. The computer-readable medium according to claim 19, wherein the program comprises further instructions for adding at least one of a differential coefficient and an acceleration coefficient to the extracted parameter.