METHOD AND APPARATUS FOR VOICE ACTIVITY DETECTION

• US 20120022863A1
• Filed: 05/03/2011
• Published: 01/26/2012
• Est. Priority Date: 07/21/2010
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: CHO, Nam-gook, KIM, Eun-kyoung
• Granted Patent: US 8,762,144 B2
• Time in Patent Office: Days
• US Class Current: 704/233
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence
• CPC Class Codes: G10L 25/03 characterised by the type of extracted parameters
• CPC Class Codes: G10L 25/84 for discriminating voice from noise

Abstract

A method and apparatus for detecting voice activity are disclosed. The method of detecting voice activity is performed in a Continuous Listening environment which includes: extracting a feature parameter from a frame signal; determining whether the frame signal is a voice signal or a noise signal by comparing the feature parameter with model parameters of a plurality of comparison signals, respectively; and outputting the frame signal when the frame signal is determined to be a voice signal. The apparatus includes a classifier module which extracts a feature parameter from a frame signal, and generating labeling information with respect to the frame signal by comparing the feature parameter with model parameters of a plurality of comparison signals; and

a voice detection unit which determines whether the frame signal is a noise signal or a voice signal with reference to the labeling information, and outputting the frame signal when the frame signal is determined to be a voice signal.

Claims

1. A method of detecting voice activity, the method being performed in a Continuous Listening environment comprising:

extracting a feature parameter from a frame signal;

determining whether the frame signal is a voice signal or a noise signal by comparing the feature parameter with model parameters of a plurality of comparison signals, respectively; and

outputting the frame signal as a voice signal when the frame signal is determined to be a voice signal.

2. The method of claim 1, further comprising calculating an energy value of the frame signal,

wherein the extracting a feature parameter comprises extracting a feature parameter from the frame signal when the energy value of the frame signal is equal to or greater than a reference value.

3. The method of claim 2, further comprising determining the frame signal to be a silence signal when the energy value of the frame signal has been determined to be smaller than the reference value, and not outputting a frame signal that has been determined to be a silence signal.

4. The method of claim 1, wherein the extracting a feature parameter comprises extracting from the frame signal at least one of a short-time energy feature, a zero-crossing rate, spectral centroid, spectral rolloff, spectral flux, spectral flatness, and a Mel-frequency ceptral coefficient.

5. The method of claim 1, wherein the comparison signals comprise first through n-th noise signals, where n is a natural number equal to or greater than 2, and a voice signal,

wherein the determining whether the frame signal is a voice signal or a noise signal comprises calculating likelihood values M1 through Mn denoting a degree of similarity between parameters of modeled characteristics of a first noise signal through the n-th noise signal and the feature parameter, and calculating a likelihood value Mn+1 indicating a degree of similarity between a parameter of modeled characteristics of the voice signal and the feature parameter.

6. The method of claim 5, wherein the determining whether the frame signal is a voice signal or a noise signal comprises calculating a ratio of a difference value between Mn+1 and a second greatest likelihood value, to a difference value between the second greatest likelihood value and a third greatest likelihood value when Mn+1 is the greatest among the likelihood values M1 through Mn+1, generating a first voice signal displaying labeling information with respect to the frame signal, when the calculated ratio is greater than a critical value, and generating a second voice signal displaying labeling information with respect to the frame signal when the calculated ratio is equal to or smaller than the critical value.

7. The method of claim 5, wherein the determining whether the frame signal is a voice signal or a noise signal comprises, when the greatest value among M1 through Mn+1 is Mk (k is a natural number that is equal to or smaller than n), generating k-th noise signal displaying labeling information with respect to the frame signal.

8. The method of claim 6, wherein the determining whether the frame signal is a voice signal or a noise signal comprises generating an index value indicating whether the frame signal is a noise signal, a voice signal, or an uncertain voice signal, by using the labeling information.

9. The method of claim 8, wherein the determining whether the frame signal is a voice signal or a noise signal comprises correcting the index value with respect to the frame signal with reference to index values, with respect to at least one of previous and subsequent frame signals of the frame signal, when the index value of the frame signal indicates that the frame signal is an uncertain voice signal.

10. The method of claim 9, wherein the outputting the frame signal comprises outputting the frame signal when the index value indicates that the frame signal is a voice signal.

11. The method of claim 9, wherein the outputting the frame signal comprises detecting a start point and an end point of a voice with reference to the index value of the frame signal.

12. An apparatus for detecting voice activity, the apparatus operating in a Continuous Listening environment and comprising:

a classifier module which extracts a feature parameter from a frame signal, and generating labeling information with respect to the frame signal by comparing the feature parameter with model parameters of a plurality of comparison signals; and

a voice detection unit which determines whether the frame signal is a noise signal or a voice signal with reference to the labeling information, and outputting the frame signal when the frame signal is determined to be a voice signal.

13. The apparatus of claim 12, further comprising an energy-based sound detector which calculates an energy value of the frame signal,

wherein the energy-based sound detector transmits the frame signal to the classifier module and the voice detection unit when the energy value of the frame signal is equal to or greater than a reference value.

14. The apparatus of claim 13, wherein when the energy value is smaller than the reference value, the energy-based sound detector generates information indicating that the frame signal is determined to be a silence signal, and transmits the generated information to the voice detection unit.

15. The apparatus of claim 12, wherein the classifier module extracts from the frame signal at least one of a short-time energy feature, a zero-crossing rate, spectral centroid, spectral rolloff, spectral flux, spectral flatness, and a Mel-frequency ceptral coefficient.

16. The apparatus of claim 12, wherein the comparison signals comprise first through n-th noise signals, where n is a natural number equal to or greater than 2, and a voice signal,

wherein the classifier module calculates likelihood values M1 through Mn denoting a degree of similarity between parameters of modeled characteristics of the first noise signal through the n-th noise signal and the feature parameter, and calculating a likelihood value Mn+1 indicating a degree of similarity between a parameter of modeled characteristics of the voice signal and the feature parameter.

17. The apparatus of claim 16, wherein the classifier module calculates a ratio of a difference value between the Mn+1 and a second greatest likelihood value to a difference value between the second greatest likelihood value and a third greatest likelihood value when Mn+1 is the greatest among the likelihood values M1 through Mn+1, the classifier module generating a first voice signal displaying labeling information with respect to the frame signal when the ratio is greater than a critical value, and the classifier module generating a second voice signal displaying labeling information with respect to the frame signal, when the ratio is equal to or smaller than the critical value.

18. The apparatus of claim 16, wherein the classifier module generates, when the greatest value among M1 through Mn+1 is Mk (k is a natural number that is equal to or smaller than n), displaying labeling information for the k-th noise signal with respect to the frame signal.

19. The apparatus of claim 17, wherein the voice detection unit generates an index value indicating whether the frame signal is a noise signal, a voice signal, or an uncertain voice signal, by using the labeling information.

20. The apparatus of claim 19, wherein the voice detection unit corrects the index value with respect to the frame signal, with reference to index values of at least one of previous and subsequent frame signals of the frame signal when the index value indicates that the frame signal is an uncertain voice signal.

21. The apparatus of claim 20, wherein the voice detection unit outputs the frame signal when the index value indicates that the frame signal is a voice signal.

22. The apparatus of claim 20, wherein the voice detection unit detects a starting point and an end point of a voice with reference to the index value.

23. A computer readable recording medium having embodied thereon a program for executing a method of detecting voice activity, wherein the program, when executed on a processor of a computer, causes the computer to implement the method in a Continuous Listening environment, the method comprising:

extracting a feature parameter from a frame signal;

determining whether the frame signal is a voice signal or a noise signal by comparing the feature parameters with model parameter of a plurality of comparison signals, respectively; and

outputting the frame signal when the frame signal is determined to be a voice signal.