Sound source classification

• US 7,117,149 B1
• Filed: 08/30/1999
• Issued: 10/03/2006
• Est. Priority Date: 08/30/1999
• Status: Expired due to Term

• Current Assignee: 2236008 Ontario Inc. (Blackberry Limited)
• Original Assignee: Harman Becker Automotive Systems GmbH (Samsung Electronics Co. Ltd.)
• Inventors: Zakarauskas, Pierre
• Primary Examiner(s): Armstrong, Angela
• Application Number:
  US09/385,975
• Time in Patent Office: 2,591 Days
• Field of Search: 704/233, 704/241, 704/243, 704/201, 704/205, 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 and method to identify a sound source among a group of sound sources. The invention matches the acoustic input to a number of signal models, one per source class, and produces a goodness-of-match number for each signal model. The sound source is declared to be of the same class as that of the signal model with the best goodness-of-match if that score is sufficiently high. The data are recorded with a microphone, digitized and transformed into the frequency domain. A signal detector is applied to the transient. A harmonic detection method can be used to determine if the sound source has harmonic characteristics. If at least some part of a transient contains signal of interest, the spectrum of the signal after resealing is compared to a set of signal models, and the input signal'"'"'s parameters are fitted to the data. The average distortion is calculated to compare patterns with those of sources that used in training the signal models. Before classification can occur, a source model is trained with signal data. Each signal model is built by creating templates from input signal spectrograms when they are significantly different from existing templates. If an existing template is found that resembles the input pattern, the template is averaged with the pattern in such a way that the resulting template is the average of all the spectra that matched that template in the past.

Claims

1. A method for classifying acoustic signal within a digitized acoustic input signal, including:

(a) transforming the digitized acoustic input signal to a time-frequency representation;

(b) isolating transient sounds within the time-frequency representation;

(c) estimating background noise and including long transients without signal content and background noise between transients in such estimating;

(d) detecting the presence of harmonics in the time-frequency representation;

(e) rescaling the time-frequency representation;

(f) comparing the rescaled time-frequency representation of each transient sound containing any signal of interest with a plurality of time-frequency spectrogram templates in a first signal model, in response to detecting the presence of harmonics, and determining a matching spectrogram template that best matches the rescaled time-frequency representation, based in part on distance between the rescaled time-frequency representation and the plurality of time-frequency spectrogram templates; and

(g) assigning the digitized acoustic input signal to the first signal model based on the matching spectrogram template.

2. The method of claim 1, where the step of assigning further includes the step of rejecting a second signal model that is does not meet a selected score threshold.

3. A system for classifying acoustic signal within a digitized acoustic input signal, including:

(a) computational means for transforming the digitized acoustic input signal to a time-frequency representation;

(b) computational means for isolating transient sounds within the time-frequency representation;

(c) computational means for estimating background noise and including long transients without signal content and background noise between transients in such estimating;

(d) computational means for detecting the presence of harmonics in the time-frequency representation;

(e) computational means for resealing the time-frequency representation;

(f) computational means for comparing the rescaled time-frequency representation of each transient sound containing any signal of interest with a plurality of time-frequency spectrogram templates in a first signal model, in response to detecting the presence of harmonics, and determining a matching spectrogram template that best matches the rescaled time-frequency representation, based in part on distance between the rescaled time-frequency representation and the plurality of time-frequency spectrogram templates; and

(g) computational means for assigning the digitized acoustic input signal to the first signal model based on the matching spectrogram template.

4. The system of claim 3, further including computational means for rejecting a second signal model that is does not meet a selected score threshold.

5. A computer program, residing on a computer-readable medium, for classifying acoustic signal within a digitized acoustic input signal, the computer program comprising instructions for causing a computer to:

(a) transform the digitized acoustic input signal to a time-frequency representation;

(b) estimate a background noise level in the time-frequency representation;

(c) for each interval of the time-frequency representation containing significant signal levels, compare the time-frequency representation of such interval, in response to estimating the background noise level to be below a pre-selected threshold, with a plurality of time-frequency spectrogram templates in a first signal model, and determine a matching spectrogram template that best matches the time-frequency representation of such interval, based in part on distance between the time-frequency representation of such interval and the plurality of time-frequency spectrogram templates; and

(d) assign the digitized acoustic input signal to the first signal model based on the matching spectrogram template.

6. The computer readable medium of claim 5, further comprising instructions for causing the computer to:

rescale the time-frequency representation according to a signal-to-noise ratio of the time-frequency representation.

7. The computer readable medium of claim 5, further comprising instructions for causing the computer to:

rescale the time-frequency representation with a weight that increases with increasing signal-to-noise ratio of the time-frequency representation.

8. A computer program, residing on a computer-readable medium, for classifying acoustic signal within a digitized acoustic input signal, the computer program comprising instructions for causing a computer to:

(a) transform the digitized acoustic input signal to a time-frequency representation;

(b) isolate transient sounds within the time-frequency representation;

(c) estimate background noise and including long transients without signal content and background noise between transients in such estimating;

(d) detect the presence of harmonics in the time-frequency representation;

(e) rescale the time-frequency representation;

(f) compare the rescaled time-frequency representation of each transient sound containing any signal of interest with a plurality of time-frequency spectrogram templates in a first signal model in response to detecting the presence of harmonics, and determine a matching spectrogram template that best matches the rescaled time-frequency representation, based in part on distance between the rescaled time-frequency representation and the plurality of time-frequency spectrogram templates; and

(g) assign the digitized acoustic input signal to the first signal model based on the matching spectrogram template.

9. The method of claim 8, further including instructions for causing a computer to reject a second signal model that is does not meet a selected score threshold.

10. An acoustic signal classifier, comprising:

an input device for receiving an acoustic signal;

a memory comprising a classification program that;

transforms the acoustic signal into an input time-frequency representation;

detects harmonics in the input time-frequency representation;

matches, in response to detected harmonics in the acoustic signal, the input time-frequency representation with a plurality of time-frequency spectrogram templates in a first signal model and in a second signal model;

determines distances between the input time-frequency representation and the time-frequency spectrogram templates in the first and second signal models; and

classifies the acoustic signal as corresponding to at least one of the first or second signal models based on the distances; and

a processor coupled to the memory for executing the classification program.

11. An acoustic signal classifier according to claim 10, where the classification program also scales the input time-frequency representation according to a weight that increases with increasing signal to noise-to-noise ratio of the input time-frequency representation.

12. An acoustic signal classifier according to claim 10, where the classification program also detects harmonic content in the input time-frequency representation and proceeds with pattern matching when harmonic content is present.

13. An acoustic signal classifier according to claim 10, where the classification program also determines a power level as a function of time and frequency for the input time-frequency representation and estimates a background nose level, and proceeds with pattern matching when the power level exceeds the background noise level by a pre-selected threshold.

14. An acoustic signal classifier according to claim 13, where the pre-selected threshold is a pre-selected number of standard deviations beyond a mean background noise level.