Robust detection and classification of objects in audio using limited training data

• US 7,263,485 B2
• Filed: 05/28/2003
• Issued: 08/28/2007
• Est. Priority Date: 05/31/2002
• Status: Expired due to Fees

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Wark, Timothy John
• Primary Examiner(s): Azad; Abul K.
• Application Number:
  US10/446,099
• Publication Number:
  US 20030231775A1
• Time in Patent Office: 1,553 Days
• Field of Search: None
• US Class Current: 704/240
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

A method (200) and apparatus (100) for classifying a homogeneous audio segment are disclosed. The homogeneous audio comprises a sequence of audio samples (x(n)). The method (200) starts by forming a sequence of frames (701-704) along the sequence of audio samples (x(n)), each frame (701-704) comprising a plurality of the audio samples (x(n)). The homogeneous audio segment is next divided (206) into a plurality of audio clips (711-714), with each audio clip being associated with a plurality of the frames (701-704). The method (200) then extracts (208) at least one frame feature for each clip (711-714). A clip feature vector (f) is next extracted from frame features of frames associated with the audio clip (711-714). Finally the segment is classified based on a continuous function during the distribution of the clip feature vectors (f).

Claims

1. A method of classifying a homogeneous audio segment into one of a plurality of classes, said method comprising the steps of:

dividing said homogeneous audio segment into a plurality of sub-segments;

extracting for each sub-segment a feature vector; and

classifying said homogeneous audio segment by comparing said feature vectors of said plurality of sub-segments with a plurality of continuous distribution functions, wherein each continuous distribution function defines one of said plurality of classes.

2. The method as claimed in claim 1 wherein said feature vector includes at least one feature extracted from the group comprising:

a bandwidth of a plurality of audio samples of said sub-segment; and

an energy value of a plurality of audio samples of said sub-segment.

3. The method as claimed in claim 1 wherein said feature vector comprises at least two features selected from the group consisting of:

Volume standard deviation;

Vlume dynamic range;

Volume undulation;

4 Hz volume peak;

Zero-crossing rate STD;

Bandwidth; and

Pitch STD.

4. The method as claimed in claim 1 wherein said classifying step comprises the sub-steps of:

calculating a measure of similarity between said feature vectors of said plurality of sub-segments and each of said continuous distribution functions associated with said plurality of classes;

determining the highest measure of similarity;

determining a confidence measure for said highest measure of similarity;

comparing said confidence measure with a confidence threshold; and

classifying said homogeneous audio segment as belonging to the class associated with said highest measure of similarity, upon said confidence measure being greater than said confidence threshold.

5. The method as claimed in claim 4 wherein said classifying step comprises the further sub-step of:

classifying said homogeneous audio segment as belonging to an unknown class upon said confidence measure being less than said confidence threshold.

6. The method as claimed in claim 4 wherein said confidence measure includes a measure of separation between said continuous distribution functions.

7. The method as claimed in claim 1 wherein said continuous distribution functions are Gaussian Mixture Models.

8. An apparatus for classifying a homogeneous audio segment into one of a plurality of classes, said apparatus comprising:

means for dividing said homogeneous audio segment into a plurality of sub-segments;

means for extracting for each sub-segment a feature vector; and

means for classifying said homogeneous audio segment by comparing said feature vectors of said plurality of sub-segments with a plurality of continuous distribution functions, wherein each continuous distribution function defines one of said plurality of classes.

9. The apparatus as claimed in claim 8 wherein said feature vector includes at least one feature extracted from the group comprising:

a bandwidth of a plurality of audio samples of said sub-segment; and

an energy value of a plurality of audio samples of said sub-segment.

10. The apparatus as claimed in claim 8 wherein said feature vector comprises at least two features selected from the group consisting of:

Volume standard deviation;

Volume dynamic range;

Volume undulation;

4 Hz volume peak;

Zero-crossing rate STD;

Bandwidth; and

Pitch STD.

11. The apparatus as claimed in claim 8 wherein said means for classifying comprises:

means for calculating a measure of similarity between said feature vectors of said plurality of sub-segments and each of said continuous distribution functions associated with said plurality of classes;

means for determining the highest measure of similarity;

means for determining a confidence measure for said highest measure of similarity;

means for comparing said confidence measure with a confidence threshold; and

means for classifying said homogeneous audio segment as belonging to the class associated with said highest measure of similarity, upon said confidence measure being greater than said confidence threshold.

12. The apparatus as claimed in claim 11 wherein said means for classifying further comprises means for classifying said homogeneous audio segment as belonging to an unknown class upon said confidence measure being less than said confidence threshold.

13. The apparatus as claimed in claim 11 wherein said confidence measure includes a measure of separation between said continuous distribution functions.

14. A program stored in a memory medium for classifying a homogeneous audio segment into one of a plurality of classes, said program comprising;

code for dividing said homogeneous audio segment into a plurality of sub-segments;

code for extracting for each sub-segment a feature vector; and

code for classifying said homogeneous audio segment by comparing said feature vectors of said plurality of sub-segments with a plurality of continuous distribution functions, wherein each continuous distribution function defines one of said plurality of classes.

15. The program as claimed in claim 14 wherein said code for classifying comprises:

code for calculating a measure of similarity between said feature vectors of said plurality of sub-segments and each of said continuous distribution functions associated with said plurality of classes;

code for determining the highest measure of similarity;

code for determining a confidence measure for said highest measure of similarity;

code for comparing said confidence measure with a confidence threshold; and

code for classifying said homogeneous audio segment as belonging to the class associated with said highest measure of similarity, upon said confidence measure being greater than said confidence threshold.

16. The program as claimed in claim 15 wherein said code for classifying further comprises:

code for classifying said homogeneous audio segment as belonging to an unknown class upon said confidence measure being less than said confidence threshold.