Acoustic signal classification system

• US 7,957,967 B2
• Filed: 09/29/2006
• Issued: 06/07/2011
• Est. Priority Date: 08/30/1999
• Status: Expired due to Fees

• Current Assignee: 2236008 Ontario Inc. (Blackberry Limited)
• Original Assignee: QNX Software Systems Company (Blackberry Limited)
• Inventors: Zakarauskas, Pierre
• Primary Examiner(s): Armstrong; Angela A
• Application Number:
  US11/540,153
• Publication Number:
  US 20070033031A1
• Time in Patent Office: 1,712 Days
• Field of Search: 704/201, 704/205, 704/233, 704/226
• US Class Current: 704/233
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• 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 classifies the source of an input signal. The system determines whether a sound source belongs to classes that may include human speech, musical instruments, machine noise, or other classes of sound sources. The system is robust, performing classification despite variation in sound level and noise masking. Additionally, the system consumes relatively few computational resources and adapts over time to provide consistently accurate classification.

Claims

1. A method for classifying an input signal, comprising:

establishing a first signal model;

obtaining an input signal;

determining a time-frequency representation of the input signal;

determining, by a processor, whether a harmonic is present in the time-frequency representation;

initiating spectrogram template matching, using the first signal model, in response to determining that the harmonic is present; and

forgoing spectrogram template matching when the harmonic is not present;

where determining whether the harmonic is present comprises;

determining a frequency range to scan;

scanning the time-frequency representation over the frequency range;

identifying local peaks in the frequency range that exceed neighboring spectrum values by more than a peak threshold;

counting votes for the local peaks;

determining a normalized vote ratio using an expected number of harmonics in the frequency range; and

comparing the normalized vote ratio to a harmonic threshold.

2. The method of claim 1, where spectrogram template matching comprises:

determining a matching spectrogram template in the first signal model for the time-frequency representation; and

assigning the input signal to the first signal model based on the matching spectrogram template.

3. The method of claim 2, further comprising:

updating the first signal model by incorporating the time-frequency representation into the matching spectrogram template.

4. The method of claim 3, where incorporating comprises:

applying a weight to the matching spectrogram template to obtain a weighted matching spectrogram template; and

averaging the weighted matching spectrogram template with the time-frequency representation.

5. The method of claim 4, where the weight is proportional to a signal-to-noise ratio of the matching spectrogram template.

6. The method of claim 1, further comprising:

updating the first signal model.

7. The method of claim 6, where updating comprises creating a new spectrogram template in the first signal model based on the time-frequency representation.

8. The method of claim 1, where spectrogram template matching comprises selecting a matching spectrogram template, by the processor, from a plurality of spectrogram templates based on an average Euclidean distance between the time-frequency representation and at least one of the plurality of spectrogram templates.

9. An acoustic signal classification system comprising:

a processor;

memory coupled to the processor, the memory comprising a first signal model and instructions operable to cause the processor to;

obtain an input signal;

determine a time-frequency representation of the input signal;

determine whether a harmonic is present in the time-frequency representation;

initiate spectrogram template matching, using the first signal model, in response to determining the harmonic is present; and

forgo spectrogram template matching when the harmonic is not present;

where the instructions that cause the processor to determine whether the harmonic is present are further operable to cause the processor to;

establish a frequency range to scan;

scan the time-frequency representation over the frequency range;

identify local peaks in the frequency range that exceed neighboring spectrum values by more than a peak threshold;

count votes for the local peaks;

determine a normalized vote ratio using an expected number of harmonics in the frequency range; and

compare the normalized vote ratio to a harmonic threshold.

10. The system according to claim 9, where the instructions are further operable to cause the processor to:

determine a matching spectrogram template in the first signal model for the time-frequency representation; and

assign the acoustic input signal to the first signal model based on the matching spectrogram template.

11. The system according to claim 10, where the instructions are further operable to cause the processor to update the first signal model by incorporating the time-frequency representation into the matching spectrogram template.

12. The system according to claim 11, where the instructions are further operable to cause the processor to:

generate a weighted matching spectrogram template;

add the time-frequency representation to the weighted matching spectrogram template to obtain a sum; and

average the sum to obtain an adjusted template.

13. The system according to claim 12, where the instructions that cause the processor to generate the weighted matching spectrogram template are further operable to cause the processor to apply a weight to the matching spectrogram template, the weight proportional to a signal-to-noise ratio of the matching spectrogram template.

14. The system according to claim 9, where the instructions are further operable to cause the processor to update the first signal model.

15. The system according to claim 14, where the instructions are further operable to cause the processor to update the first signal model by creating a new spectrogram template in the signal model based on the time-frequency representation.

16. A method for classifying an input signal, comprising:

establishing a first signal model;

obtaining an input signal;

determining a time-frequency representation of the input signal;

determining, by a processor, whether a harmonic is present in the time-frequency representation;

initiating spectrogram template matching, using the first signal model, in response to determining that the harmonic is present; and

forgoing spectrogram template matching when the harmonic is not present;

where spectrogram template matching comprises;

comparing the time-frequency representation with a plurality of time-frequency spectrogram templates;

calculating a first root mean square distance difference between the time-frequency representation and a first template of the plurality of time-frequency spectrogram templates;

calculating a second root mean square distance difference between the time-frequency representation and a second template of the plurality of time-frequency spectrogram templates;

identifying the first template as a matching spectrogram template in response to determining that the first root mean square distance difference is lower than the second root mean square distance difference; and

identifying the second template as the matching spectrogram template, by the processor, in response to a determination that the second root mean square distance difference is lower than the first root mean square distance difference.

17. A method for classifying an input signal, comprising:

obtaining an input signal;

determining a time-frequency representation of the input signal;

detecting a harmonic present in the time-frequency representation; and

initiating spectrogram template matching by a processor in response to detecting the harmonic;

where detecting the harmonic comprises;

determining a frequency range to scan;

scanning the time-frequency representation over the frequency range;

identifying local peaks in the frequency range that exceed neighboring spectrum values by more than a peak threshold;

counting votes for the local peaks;

determining a normalized vote ratio using an expected number of harmonics in the frequency range; and

comparing the normalized vote ratio to a harmonic threshold.

18. A product comprising:

a machine readable medium; and

instructions stored on the machine readable medium that cause a processor in a signal processing system to;

establish a first signal model;

obtain an input signal;

determine a time-frequency representation of the input signal;

determine whether a harmonic is present in the time-frequency representation;

initiate spectrogram template matching, using the first signal model, in response to determining that the harmonic is present; and

forgo spectrogram template matching when the harmonic is not present;

where the instructions that cause the system to determine whether the harmonic is present are further operable to cause the processor to;

establish a frequency range to scan;

scan the time-frequency representation over the frequency range;

identify local peaks in the frequency range that exceed neighboring spectrum values by more than a peak threshold;

count votes for the local peaks;

determine a normalized vote ratio using an expected number of harmonics in the frequency range; and

compare the normalized vote ratio to a harmonic threshold.

19. The product of claim 18, where:

the spectrogram template matching searches the first signal model; and

the instructions are further operable to cause the system to update the first signal model by creating a new spectrogram template in the first signal model based on the time-frequency representation.

20. The product of claim 19, where:

the spectrogram template matching results in a matching spectrogram template in the first signal model; and

where the instructions are further operable to cause the system to update the first signal model by incorporating the time-frequency representation into the matching spectrogram template.

21. The product of claim 18, where the instructions are further operable to cause the processor to:

generate a weighted matching spectrogram template;

add the time-frequency representation to the weighted matching spectrogram template to obtain a sum; and

average the sum to obtain an adjusted template.