Intervalgram representation of audio for melody recognition
A system, method, and computer readable storage medium generates an audio fingerprint for an input audio clip that is robust to differences in key, instrumentation, and other performance variations. The audio fingerprint includes a sequence of intervalgrams that represent a melody in an audio clip according pitch intervals between different time points in the audio clip. The fingerprint for an input audio clip can be compared to a set of reference fingerprints in a reference database to determine a matching reference audio clip.
1. A computer-implemented method for matching audio clips, the method executed by a computer system, and comprising:
receiving an audio chromagram representing an input audio clip, the audio chromagram comprising a sequence of vectors, each vector in the audio chromagram corresponding to a different time point of the input audio clip, and each vector representing a distribution of audio pitches at the corresponding time point of the input audio clip;
selecting a sampling of different reference time points within the audio chromagram;
for each of the selected reference time points, generating a chroma block having a plurality of vectors, each of the plurality of vectors in the chroma block corresponding to a different time sub-period of the input audio clip, and each vector representing a weighted average of distributions of audio pitches within the corresponding time sub-period;
for each of the selected reference time points, generating a reference vector representing a reference distribution of pitches for the selected reference time point in the audio chromagram;
applying a circular cross-correlation of the vectors of each chroma block against the reference vector to produce a sequence of intervalgram blocks for the input audio clip, wherein the sequence of intervalgram blocks comprises an intervalgram representation for the input audio clip, each intervalgram block associated with a different time period within the input audio clip, and each intervalgram block representing a distribution of pitch intervals occurring between different sub-periods within the time period;
comparing the intervalgram representation for the input audio clip to stored intervalgram representations corresponding to reference audio clips in a reference database;
selecting a reference audio clip from the reference database having an intervalgram representation best matching the intervalgram representation for the input audio clip; and
generating a recognition result indicative of the selected reference audio clip.
2. The method of claim 1 , wherein generating the chroma block comprises:
extracting a block of chroma vectors from the chromagram, the block of chroma vectors corresponding to one of the different time periods within the input audio clip; and
averaging the chroma vectors in the block of chroma vectors across time into a number of time bins, each time bin corresponding to a different time sub-period in the chroma block.
3. The method of claim 2 , wherein averaging the chroma vectors in the block of chroma vectors comprises applying a non-uniform sub-sampling such that time bins near the reference time are narrower than time bins at the edges of the block of chroma vectors.
4. The method of claim 1 , wherein generating the reference vector comprises computing a weighted average of a plurality of chroma vectors within a limited time range surrounding the reference time.
5. The method of claim 4 , wherein computing the weighted average of the plurality of chroma vectors comprises applying a triangle weighting function to the plurality of chroma vectors.
6. The method of claim 1 , further comprising generating the audio chromagram, wherein generating the audio chromagram comprises:
receiving a pitchogram comprising a sequence of vectors, each vector corresponding to a different time point in the input audio clip, and each vector in the pitchogram representing a distribution of pitches at the time point in the input audio clip; and
quantizing and wrapping the distribution of pitches in of the sequence of vectors in the pitchogram into a number of chroma bins according to a weighted smoothing function, wherein each chroma bin represents a pitch independent of octave.
7. The method of claim 1 , further comprising storing the intervalgram representation for the input audio clip to the reference database as an additional reference fingerprint.
8. A computer-implemented method for generating a reference database of audio fingerprints representative of melodies in a corresponding set of reference audio clips, the method executed by a computer system, and comprising:
receiving a reference audio clip;
generating an audio chromagram representing the reference audio clip, the audio chromagram comprising a sequence of vectors, each vector in them audio chromagram corresponding to a different time point of the reference audio clip, and each vector representing a distribution of audio pitches at the corresponding time point of the reference audio clip;
selecting a set of different reference time points within the audio chromagram;
for each of the selected reference time points, generating a chroma block having a plurality of vectors, each of the plurality of vectors in the chroma block corresponding to a different time sub-period of the reference audio clip, and each vector representing a weighted average of distributions of audio pitches within the corresponding time sub-period;
for each of the selected reference time points, generating a reference vector representing a reference distribution of pitches for the selected reference time point in the audio chromagram;
applying a circular cross-correlation of the vectors of each chroma block against the reference vector to produce a sequence of intervalgram blocks for the reference audio clip, wherein the sequence of intervalgram blocks comprises an intervalgram representation for the reference audio clip, each intervalgram block associated with a different time period within the reference audio clip, and each intervalgram block representing a distribution of pitch intervals occurring between different sub-periods within the time period; and
storing the intervalgram representation as a reference fingerprint in the reference database.
9. The method of claim 8 , wherein storing the intervalgram representation as the reference fingerprint in the reference database comprises:
indexing the reference fingerprint according to features of the reference fingerprint.
10. A non-transitory computer readable storage medium storing computer-executable program instructions for matching audio clips, the program instructions when executed cause a processor to perform steps of:
receiving an audio chromagram representing an input audio clip, the audio chromagram comprising a sequence of vectors, each vector in the audio chromagram corresponding to a different time point of the input audio clip, and each vector representing a distribution of audio pitches at the corresponding time point of the input audio clip;
selecting a sampling of different reference time points within the audio chromagram;
for each of the selected reference time points, generating a chroma block having a plurality of vectors, each of the plurality of vectors in the chroma block corresponding to a different time sub-period of the input audio clip, and each vector representing a weighted average of distributions of audio pitches within the corresponding time sub-period;
for each of the selected reference time points, generating a reference vector representing a reference distribution of pitches for the selected reference time point in the audio chromagram;
applying a circular cross-correlation of the vectors of each chroma block against the reference vector to produce a sequence of intervalgram blocks for the input audio clip, wherein the sequence of intervalgram blocks comprises an intervalgram representation for the input audio clip, each intervalgram block associated with a different time period within the input audio clip, and each intervalgram block representing a distribution of pitch intervals occurring between different sub-periods within the time period;
comparing the intervalgram representation for the input audio clip to stored intervalgram representations corresponding to reference audio clips in a reference database;
selecting a reference audio clip from the reference database having an intervalgram representation best matching the intervalgram representation for the input audio clip; and
generating a recognition result indicative of the selected reference audio clip.
11. The computer-readable storage medium of claim 10 , wherein generating the chroma block comprises:
extracting a block of chroma vectors from the chromagram, the block of chroma vectors corresponding to one of the different time periods within the input audio clip; and
averaging the chroma vectors in the block of chroma vectors across time into a number of time bins, each time bin corresponding to a different time sub-period in the chroma block.
12. The computer-readable storage medium of claim 11 , wherein averaging the chroma vectors in the block of chroma vectors comprises applying a non-uniform sub-sampling such that time bins near the reference time are narrower than time bins at the edges of the block of chroma vectors.
13. The computer-readable storage medium of claim 10 , wherein generating the reference vector comprises computing a weighted average of a plurality of chroma vectors within a limited time range surrounding the reference time.
14. The computer-readable storage medium of claim 13 , wherein computing the weighted average of the plurality of chroma vectors comprises applying a triangle weighting function to the plurality of chroma vectors.
15. The computer-readable storage medium of claim 10 , further comprising program instructions for generating the audio chromagram, wherein generating the audio chromagram comprises:
receiving a pitchogram comprising a sequence of vectors, each of the vectors corresponding to a different time point in the input audio clip, and each vector in the pitchogram representing a distribution of pitches at the time point in the input audio clip; and
quantizing and wrapping the distribution of pitches in of the sequence of vectors in the pitchogram into a number of chroma bins according to a weighted smoothing function, wherein each chroma bin represents a pitch independent of octave.
16. The computer-readable storage medium 10 , further comprising storing the intervalgram representation for the input audio clip to the reference database as an additional reference fingerprint.
17. A non-transitory computer readable storage medium storing computer-executable program instructions for generating a reference database of audio fingerprints representative of melodies in a corresponding set of reference audio clip, the program instructions when executed cause a processor to perform steps of:
receiving a reference audio clip;
generating an audio chromagram representing the reference audio clip, the audio chromagram comprising a sequence of vectors, each vector in them audio chromagram corresponding to a different time point of the reference audio clip, and each vector representing a distribution of audio pitches at the corresponding time point of the reference audio clip;
selecting a set of different reference time points within the audio chromagram;
for each of the selected reference time points, generating a chroma block having a plurality of vectors, each of the plurality of vectors in the chroma block corresponding to a different time sub-period of the reference audio clip, and each vector representing a weighted average of distributions of audio pitches within the corresponding time sub-period;
for each of the selected reference time points, generating a reference vector representing a reference distribution of pitches for the selected reference time point in the audio chromagram;
applying a circular cross-correlation of the vectors of each chroma block against the reference vector to produce a sequence of intervalgram blocks for the reference audio clip, wherein the sequence of intervalgram blocks comprises an intervalgram representation for the reference audio clip, each intervalgram block associated with a different time period within the reference audio clip, and each intervalgram block representing a distribution of pitch intervals occurring between different sub-periods within the time period; and
storing the intervalgram representation as a reference fingerprint in the reference database.
18. The computer-readable storage medium of claim 17 , wherein storing the intervalgram representation as the reference fingerprint in the reference database comprises:
indexing the reference fingerprint according to features of the reference fingerprint.