Selective sampling for sound signal classification

• US 20050043957A1
• Filed: 08/21/2003
• Published: 02/24/2005
• Est. Priority Date: 08/21/2003
• Status: Active Grant

• Current Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Original Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Inventors: Lin, Xiaofan
• Granted Patent: US 7,340,398 B2
• Time in Patent Office: Days
• US Class Current: 704/277
• CPC Class Codes: G10L 17/26 Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices

Abstract

A system and method of selective sampling for sound signal classification is disclosed. The method of the present invention discloses the elements of: receiving a sound signal; specifying meta-data to be extracted from the sound signal; dividing the sound signal into a set of frames; applying a fitness function to the frames to create a set of fitness data; selecting a frame from the set of frames, if the frame'"'"'s corresponding fitness datum within the set of fitness data exceeds a predetermined threshold value; extracting the meta-data from the selected frames; and classifying the sound signal based on the meta-data extracted from the selected frames. The system of the present invention discloses means for implementing the method.

Claims

1. A method for sound signal classification, comprising:

receiving a sound signal;

specifying meta-data to be extracted from the sound signal;

dividing the sound signal into a set of frames;

applying a fitness function to the frames to create a set of fitness data;

selecting a frame from the set of frames, if the frame'"'"'s corresponding fitness datum within the set of fitness data exceeds a predetermined threshold value;

extracting the meta-data from the selected frames; and

classifying the sound signal based on the meta-data extracted from the selected frames.

2. The method of claim 1:

wherein the sound signal is a speech signal.

3. The method of claim 1 wherein specifying includes:

specifying age range meta-data.

4. The method of claim 1 wherein specifying includes:

specifying gender meta-data.

5. The method of claim 4 wherein selecting includes:

setting the threshold so that a ratio of frames selected to frames not selected is between about 1;

2 and about 1;

3.

6. The method of claim 1 wherein specifying includes:

specifying accent meta-data.

8. The method of claim 1 wherein specifying includes:

specifying identity meta-data.

9. The method of claim 1 wherein dividing includes:

dividing the sound signal into a set of time frames.

10. The method of claim 1 wherein dividing includes:

dividing the sound signal into a set of equal length time frames.

11. The method of claim 1 wherein applying includes:

calculating a signal strength of the sound signal frame.

12. The method of claim 1 wherein selecting includes:

selecting a frame for meta-data extraction, if the frame'"'"'s fitness datum exceeds a greatest fitness datum within the set of fitness data by a predetermined margin.

13. The method of claim 1 wherein extracting includes:

extracting the meta-data from the selected frames using a Multi-Layer Perceptron (MLP) neural network.

14. The method of claim 13 wherein extracting includes:

extracting the meta-data from the selected frames using a MLP neural network having an input layer with nodes corresponding to the sound signal'"'"'s Mel-Cepstral components.

15. The method of claim 1 further wherein classifying includes:

assigning the sound signal to that meta-data class to which a largest number of the selected frames have been assigned.

17. The method of claim 1 further wherein classifying includes:

assigning the sound signal to that meta-data class having a statistically longest run-length.

7. The method of claim I wherein specifying includes:

specifying dialect meta-data.

16. The method of claim I further wherein classifying includes:

adding together each of the selected frame'"'"'s confidence scores for each meta-data class; and

assigning the sound signal to that meta-data class having a highest total confidence score.

18. A method for sound signal classification, comprising:

receiving a speech signal;

specifying meta-data to be extracted from the sound signal;

dividing the sound signal into a set of equal length time frames;

applying a fitness function to the frames to create a set of fitness data;

selecting a frame for meta-data extraction, if the frame'"'"'s fitness datum exceeds a greatest fitness datum within the set of fitness data by a predetermined margin;

extracting the meta-data from the selected frames using a Multi-Layer Perceptron (MLP) neural network;

adding together each of the selected frame'"'"'s confidence scores for each meta-data class; and

assigning the sound signal to that meta-data class having a highest total confidence score.

19. A system for sound signal classification comprising a:

means for receiving a sound signal;

means for specifying meta-data to be extracted from the sound signal;

means for dividing the sound signal into a set of frames;

means for applying a fitness function to the frames to create a set of fitness data;

means for selecting a frame from the set of frames, if the frame'"'"'s corresponding fitness datum within the set of fitness data exceeds a predetermined threshold value;

means for extracting the meta-data from the selected frames; and

means for classifying the sound signal based on the meta-data extracted from the selected frames.