Extraction and matching of characteristic fingerprints from audio signals
View Patent ↗An audio fingerprint is extracted from an audio sample, where the fingerprint contains information that is characteristic of the content in the sample. The fingerprint may be generated by computing an energy spectrum for the audio sample, resampling the energy spectrum logarithmically in the time dimension, transforming the resampled energy spectrum to produce a series of feature vectors, and computing the fingerprint using differential coding of the feature vectors. The generated fingerprint can be compared to a set of reference fingerprints in a database to identify the original audio content.
1. A method for extracting an audio fingerprint from an audio frame, the method comprising:
filtering the audio frame into a plurality of frequency bands to produce a corresponding plurality of filtered audio signals;
resampling the filtered audio signals at a nonlinear timescale;
transforming the resampled audio signals for each frequency band to produce a feature vector for the frequency band; and
computing the audio fingerprint based on the set of feature vectors.
2. The method of claim 1 , wherein filtering the audio frame into a plurality of frequency bands comprises band pass filtering the audio frame in each of the plurality of frequency bands.
3. The method of claim 1 , wherein filtering the audio frame into a plurality of frequency bands comprises performing a Fast Fourier Transform (FFT) on the audio sample.
4. The method of claim 1 , wherein the audio frame is part of an audio file stored in an MP3 format, and the filtered audio signals are obtained from an MP3 hybrid filterbank associated with the audio file.
5. The method of claim 1 , wherein the filtered audio signals are resampled at a logarithmic timescale.
6. The method of claim 1 , wherein the frequency bands are spaced linearly in a frequency axis.
7. The method of claim 1 , wherein the frequency bands overlap.
8. The method of claim 1 , wherein transforming the resampled filtered audio signal of a particular frequency band comprises performing a Fast Fourier Transform (FFT) on the resampled audio signal.
9. The method of claim 1 , wherein computing the audio fingerprint comprises differentially encoding the feature vectors for the frequency bands.
10. The method of claim 1 , further comprising:
computing an index value for the audio fingerprint, the index value comprising a portion of the audio fingerprint.
11. The method of claim 10 , wherein the index value comprises a portion of the audio fingerprint that corresponds to a set of low frequency components of the transformed audio signals.
12. The method of claim 1 , further comprising:
disregarding a portion of the audio fingerprint, where the disregarded portion of the audio fingerprint corresponds to a frequency range determined to be insignificant according to an acoustic model.
13. The method of claim 12 , wherein the acoustic model is a psychoacoustic model.
14. The method of claim 12 , wherein the acoustic model mimics the properties of an audio encoding process.
15. The method of claim 12 , wherein the acoustic model mimics the properties of an environment.
16. The method of claim 12 , wherein the acoustic model mimics the properties of an audio signal.
17. A method for creating a database of audio fingerprints, the method comprising:
generating a plurality of audio fingerprints, each audio fingerprint generating according to the method of claim 1 ;
computing an index value for each audio fingerprint, the index value comprising a portion of the audio fingerprint; and
grouping the audio fingerprints in the database according to their index values.
18. A method for detecting an edge of an audio fragment, the method comprising:
receiving a input signal representing the audio fragment;
storing in a buffer of finite length a most recent portion of the received input signal;
for each a plurality of times while receiving the input signal, computing a fingerprint for the portion of the received input signal stored in the buffer according to the method of claim 1 ;
computing a bit error rate (BER) between each of the computed fingerprints and a reference fingerprint associated with the audio fragment; and
identifying an edge of the audio fragment based on the computed BERs over time.
19. The method of claim 18 , wherein identifying an edge of the audio fragment comprises:
finding a best fit line matching a portion of the computed BERs during a period of decreasing BER over time; and
identifying a beginning edge of the audio fragment as an intersection between the best fit line and a horizontal line in which BER equal to 50%.
20. The method of claim 18 , wherein identifying an edge of the audio fragment comprises:
finding a best fit line matching a portion of the computed BERs during a period of increasing BER over time; and
identifying an ending edge of the audio fragment as an intersection between the best fit line and a horizontal line in which BER equal to 50%, subtracted by the length of the buffer.
21. The method of claim 1 , wherein transforming the resampled filtered audio signal of a particular frequency band comprises performing a Discrete Cosine Transform (DCT) on the resampled audio signal.
22. The method of claim 1 , wherein the frequency bands have a logarithmic mid-frequency distribution in a frequency axis.
23. The method of claim 1 , wherein the filtered audio signals are resampled at an exponential timescale.
24. The method of claim 1 , wherein computing the audio fingerprint comprises encoding the feature vectors for the frequency bands by assigning bit values according to a codebook table.
25. The method of claim 24 , further comprising:
training the codebook table to determine a set of codebook values; and
tuning the codebook table based on the determined set of codebook values.
26. A method for extracting an audio fingerprint from an audio frame, the method comprising:
a step for computing a spectrogram for the audio frame;
sampling the spectrogram at a nonlinear time scale for a plurality of frequency bands in the spectrogram;
a step for extracting a long-term feature vector using the samples from each of the sampled frequency bands; and
a step for generating the audio fingerprint based on the feature vectors.
27. The method of claim 26 , wherein the spectrogram is sampled at a logarithmic timescale.
28. A computer program product for extracting an audio fingerprint from an audio frame, the computer program product comprising a computer-readable medium containing computer program code for performing the method comprising:
filtering the audio frame into a plurality of frequency bands to produce a corresponding plurality of filtered audio signals;
resampling the filtered audio signals at a nonlinear timescale;
transforming the resampled audio signals for each frequency band to produce a feature vector for the frequency band; and
computing the audio fingerprint based on the set of feature vectors.
29. The computer program product of claim 28 , wherein filtering the audio frame into a plurality of frequency bands comprises band pass filtering the audio frame in each of the plurality of frequency bands.
30. The computer program product of claim 28 , wherein filtering the audio frame into a plurality of frequency bands comprises performing a Fast Fourier Transform (FFT) on the audio sample.
31. The computer program product of claim 28 , wherein the audio frame is part of an audio file stored in an MP3 format, and the filtered audio signals are obtained from an MP3 hybrid filterbank associated with the audio file.
32. The computer program product of claim 28 , wherein the filtered audio signals are resampled at a logarithmic timescale.
33. The computer program product of claim 28 , wherein the frequency bands are spaced linearly in a frequency axis.
34. The computer program product of claim 28 , wherein the frequency bands overlap.
35. The computer program product of claim 28 , wherein transforming the resampled filtered audio signal of a particular frequency band comprises performing a Fast Fourier Transform (FFT) on the resampled audio signal.
36. The computer program product of claim 28 , wherein computing the audio fingerprint comprises differentially encoding the feature vectors for the frequency bands.
37. The computer program product of claim 28 , further comprising:
computing an index value for the audio fingerprint, the index value comprising a portion of the audio fingerprint.
38. The computer program product of claim 37 , wherein the index value comprises a portion of the audio fingerprint that corresponds to a set of low frequency components of the transformed audio signals.
39. The computer program product of claim 28 , further comprising:
disregarding a portion of the audio fingerprint, where the disregarded portion of the audio fingerprint corresponds to a frequency range determined to be insignificant according to an acoustic model.
40. The computer program product of claim 39 , wherein the acoustic model is a psychoacoustic model.
41. The computer program product of claim 39 , wherein the acoustic model mimics the properties of an audio encoding process.
42. The computer program product of claim 28 , wherein the frequency bands have a logarithmic mid-frequency distribution in a frequency axis.
43. The computer program product of claim 39 , wherein the acoustic model mimics the properties of an environment.
44. The computer program product of claim 28 , wherein the filtered audio signals are resampled at an exponential timescale.
45. The computer program product of claim 39 , wherein the acoustic model mimics the properties of an audio signal.
46. The computer program product of claim 28 , wherein transforming the resampled filtered audio signal of a particular frequency band comprises performing a Discrete Cosine Transform (DCT) on the resampled audio signal.
47. The computer program product of claim 28 , wherein computing the audio fingerprint comprises encoding the feature vectors for the frequency bands by assigning bit values according to a codebook table.
48. The computer program product of claim 47 , further comprising:
training the codebook table to determine a set of codebook values; and
tuning the codebook table based on the determined set of codebook values.