Song similarity determination
Aspects of the technology described herein use acoustic features of a music track to capture information for a recommendation system. The recommendation can work without analyzing label data (e.g., genre, artist) or usage data for a track. For each audio track, a descriptor is generated that can be used to compare the track to other tracks. The comparisons between track descriptors result in a similarity measure that can be used to make a recommendation. In this process, the audio descriptors are used directly to form a track-to-track similarity measure between tracks. By measuring the similarity between a track that a user is known to like and an unknown track, a decision can be made whether to recommend the unknown track to the user.
1. One or more computer storage media comprising computer-executable instructions that when executed by a computing device cause the computing device to perform a method of selecting a song for a user, comprising:
receiving a digital recording of a song comprising a series of frames;
performing a scattering transform on each frame in the series of frames to form a plurality of feature vectors;
using the plurality of feature vectors for the digital recording as input, generating a multivariable density estimation for the digital recording;
calculating a similarity score between the digital recording and a different digital recording of a different song by comparing the multivariable density estimation for the digital recording to a second multivariable density estimation for the different digital recording; and
outputting to the user a recommendation for the song based on the similarity score being within a threshold similarity to the different digital recording and a usage data indicating the user likes the different digital recording.
2. The media of claim 1 , wherein consecutive frames within the series of frames overlap.
3. The media of claim 1 , wherein the song is performed by a first artist, the method further comprising determining that the first artist is similar to a second artist by comparing similarity scores between the song and a plurality of songs by the second artist.
4. The media of claim 1 , wherein the plurality of feature vectors are generated using a dimensional reduction method to reduce an amount of variables in an individual feature vector.
5. The media of claim 1 , wherein the multivariable density estimation is a multivariable vector generated by calculating differences between the digital recording and a mean feature distribution generated from a plurality of songs.
6. The media of claim 1 , wherein the similarity score is calculated using cosine similarity.
7. The media of claim 1 , wherein the usage data indicating the user likes the different digital recording is inclusion of the different song in a playlist.
8. A method of selecting a song for a user, the method comprising:
determining that a digital recording of a song does not have a statistically significant amount of usage data;
performing a scattering transform on frames in the digital recording to form a plurality of feature vectors;
using the plurality of feature vectors for the digital recording as input, generating a multivariable density estimation for the digital recording;
calculating a similarity score between the digital recording and a different digital recording of a different song by comparing the multivariable density estimation for the digital recording to a second multivariable density estimation for the different digital recording; and
outputting a recommendation to the user for the song based on the similarity score and a usage pattern of the user for the plurality of different songs.
9. The method of claim 8 , wherein consecutive frames in the series of frames are overlapping.
10. The method of claim 8 , wherein the recommendation is a genre classification.
11. The method of claim 8 , wherein the song is performed by a first artist, the method further comprising determining that the first artist is similar to a second artist by comparing similarity scores for the song with a plurality of songs by the second artist.
12. The method of claim 8 , wherein the song is included on a first album, the method further comprising determining that the first album is similar to a second album by comparing similarity scores for the song with a plurality of songs on the second album.
13. The method of claim 8 , wherein the recommendation is a notification suggesting the user listen to the song.
14. The method of claim 8 , wherein the usage data indicating the user likes the different digital recording is inclusion of the different song in a playlist.
15. The method of claim 8 , wherein the usage pattern for the plurality of different songs is generated by a digital music streaming service.
16. A method of selecting a song for a user comprising:
receiving a digital recording of a song that is not associated with a statistically significant amount of usage data;
performing a scattering transform on frames in the digital recording to form a plurality of feature vectors;
using the plurality of feature vectors for the digital recording as input, generating a multivariable density estimation for the digital recording;
calculating a similarity score between the digital recording and a different digital recording of a different song by comparing the multivariable density estimation for the digital recording to a second multivariable density estimation for the different digital recording; and
outputting a recommendation based on the similarity score.
17. The method of claim 16 , wherein the similarity score is calculated without usage data for the digital recording or the different digital recording.
18. The method of claim 16 , wherein the similarity score is calculated without using a genre classification, an album title, or an artist for the digital recording, or a genre classification, an album title, or an artist for the different digital recording.
19. The method of claim 16 , wherein the multivariable density estimation is a multivariable vector generated by calculating differences between the digital recording and a mean feature distribution generated from a plurality of songs.
20. The method of claim 16 , wherein the recommendation is inclusion in a playlist that includes the different song.