Back to results
US 8,977,374 B1Granted
Geometric and acoustic joint learning
AssigneeGoogle Inc.
InventorsDouglas Eck, Jay Yagnik
Life of this patent
Estimating the term…
Abstract
Described herein are methods and system for analyzing music audio. An example method includes obtaining a music audio track, calculating acoustic features of the music audio track, calculating geometric features of the music audio track in view of the acoustic features, and determining a mood of the music audio track in view of the geometric features.
Claims39
A music-analysis tool comprising:
- a music audio track selection module that is configured to select at least one music audio track from a corpus of music audio tracks;
- an acoustic-features module that is configured to calculate acoustic features of the music audio track;
- a vector-quantization module that is configured to vector quantize the calculated acoustic features;
- a geometric-features module that is configured to calculate geometric features of the music audio track based on the vector-quantized acoustic features, wherein the vector-quantization module is further configured to vector quantize the calculated geometric features;
- a cross-product module that is configured to take a cross-product of the vector-quantized acoustic features and the vector-quantized geometric features to yield a joint acoustic-geometric feature vector for the music audio track;
- and
- a playlist-selection module that is configured to select a playlist of music audio based on the joint acoustic-geometric feature vector for the music audio track.
4 more elements
A music-analysis tool comprising:
- a music audio track selection module that is configured to select at least one music audio track from a corpus of music audio tracks;
- a Mel-frequency cepstrum coefficients (MFCCs) module that is configured to calculate the MFCCs from local features of the music audio track;
- a k-means module that is configured to calculate k-means for the calculated MFCCs;
- a regional-features module that is configured to calculate regional features of the music audio track based on the calculated k-means, wherein the k-means module is further configured to calculate the k-means of the calculated regional features;
- a cross-product module that is configured to take a cross-product of the k-means of the calculated MFCCs and the k-means of the calculated regional features to yield a joint local-regional feature vector for the music audio track;
- and
- an artist-selection module that is configured to select a music artist based on the joint local-regional feature vector for the music audio track.
4 more elements
A music-analysis tool comprising:
- a music audio track selection module that is configured to select at least one music audio track from a corpus of music audio tracks;
- a Stabilized Auditory Image (SAI) module that is configured to calculate short timescale features of the music audio track;
- a winner-take-all (WTA) module that is configured to vector quantize the calculated short timescale features;
- a long-timescale-features module that is configured to calculate long timescale features of the music audio track based on the vector-quantized short timescale features, wherein the WTA module is further configured to vector quantize the calculated long timescale features;
- a cross-product module that is configured to take a cross-product of the vector-quantized short timescale features and the vector-quantized long timescale features to yield a joint short timescale-long timescale feature vector for the music audio track;
- and
- a mood-selection module that is configured to determine a mood of the music audio track based on the joint short timescale-long timescale feature vector for the music audio track.
4 more elements
A method comprising:
- obtaining a music audio track;
- calculating acoustic features of the music audio track;
- vector quantizing the calculated acoustic features;
- creating a geometric model based on the vector-quantized acoustic features;
- using the geometric model and based on the vector-quantized acoustic features, calculating geometric features of the music audio track;
- mapping the vector-quantized acoustic features onto vector-quantized geometric features to yield joint acoustic-geometric features;
- and
- using the joint acoustic-geometric features, determining a mood of the music audio track.
5 more elements
One or more computer-readable media storing processor-executable instructions that when executed cause one or more processors to perform a method comprising:
- obtaining a music audio track;
- calculating acoustic features of the music audio track using an audio spectrogram of the music audio track, wherein the audio spectrogram includes a set of clusters of interest points in the audio spectrogram, wherein the audio spectrogram further includes a centroid for each cluster, wherein the centroids form a code;
- vector quantizing the calculated acoustic features by mapping the codes into a code book;
- creating a geometric model based on the code book;
- using the geometric model and based on the code book, calculating geometric features of the music audio track;
- mapping the codebook onto the vector-quantized geometric features to yield joint acoustic-geometric features;
- and
- using the joint acoustic-geometric features, determining a mood of the music audio track.
5 more elements
One or more computer-readable media storing processor-executable instructions that when executed cause one or more processors to perform a method comprising:
- obtaining a music audio track;
- calculating acoustic features of the music audio track;
- vector quantizing the calculated acoustic features;
- creating a geometric model based on the vector-quantized acoustic features;
- using the geometric model, vector-quantizing geometric features of the music audio track;
- mapping the vector-quantized acoustic features onto the vector-quantized geometric features to yield joint acoustic-geometric features;
- and
- using the joint acoustic-geometric features, select a playlist of music audio tracks based on the joint acoustic-geometric feature vector for the music audio track.
5 more elements
A music-analysis system comprising:
- a music-analysis tool having:
- a music audio track selection module that is configured to select at least one music audio track from a corpus of music audio tracks;
- an acoustic-features module that is configured to calculate acoustic features of the obtained music audio track;
- a vector-quantization module that is configured to vector quantize the calculated acoustic features;
- a geometric-features module that is configured to calculate geometric features of the obtained music audio track based on the vector-quantized acoustic features, wherein the vector-quantization module is further configured to vector quantize the calculated geometric features;
- and
- a cross-product module that is configured to take a cross-product of the vector-quantized acoustic features and the vector-quantized geometric features to yield a joint acoustic-geometric feature vector for the obtained music audio track;
- and
- a generative-music module that is configured to compose a music audio track that is different from the obtained music audio track using the joint-acoustic-geometric feature vector for the obtained music audio track.
6 more elements
Dependencies are inferred from the claim text.