Controllable music generation
The present disclosure describes techniques for controllable music generation. The techniques comprise extracting latent vectors from unlabelled data, the unlabelled data comprising a plurality of music note sequences, the plurality of music note sequences indicating a plurality of pieces of music; clustering the latent vectors into a plurality of classes corresponding to a plurality of music styles; generating a plurality of labelled latent vectors corresponding to the plurality of music styles, each of the plurality labelled latent vectors comprising information indicating features of a corresponding music style; and generating a first music note sequence indicating a first piece of music in a particular music style among the plurality of music styles based at least in part on a particular labelled latent vector among the plurality of labelled latent vectors, the particular labelled latent vector corresponding to the particular music style.
1. A method, comprising:
extracting latent vectors from unlabelled data by an encoder, the unlabelled data comprising a plurality of music note sequences, the plurality of music note sequences indicating a plurality of pieces of music;
clustering the latent vectors into a plurality of classes corresponding to a plurality of music styles by a machine learning model;
generating a plurality of labelled latent vectors corresponding to the plurality of music styles by the machine learning model, each of the plurality of labelled latent vectors comprising information indicating features of a corresponding music style among the plurality of music styles; and
generating, by a decoder, a first music note sequence indicating a first piece of music in a particular music style among the plurality of music styles based at least in part on a particular labelled latent vector among the plurality of labelled latent vectors, the particular labelled latent vector corresponding to the particular music style.
2. The method of claim 1 , wherein the encoder comprises an encoder of a variational autoencoder (VAE), wherein the decoder comprises a decoder of the VAE, and wherein the method further comprises:
inputting the particular labelled latent vector to the decoder of the VAE.
3. The method of claim 2 , wherein the VAE is trained using unlabelled music training data, the encoder of the VAE learns to encode the unlabelled music training data into Gaussian distributions, and the decoder of the VAE learns to decode latent vectors sampled from the Gaussian distributions into pieces of music.
4. The method of claim 1 , wherein the machine learning model comprises a Gaussian Mixture Model (GMM).
5. The method of claim 4 , wherein the GMM is trained using unlabelled latent vectors extracted by an encoder of a VAE from unlabelled music training data.
6. The method of claim 1 , further comprising:
interpolating, by a controllable music generation model, a specific latent vector corresponding to a type of chord sequence with another latent vector corresponding to a random music sequence to generate a mixed latent vector corresponding to a new music sequence, wherein the specific latent vector is among the plurality of labelled latent vectors, and wherein the new music sequence is harmonic with both the chord sequence and the random music sequence.
7. The method of claim 6 , further comprising:
inputting the mixed latent vector to a VAE comprising the decoder; and
generating, by the decoder, a second music note sequence indicating a second piece of music in a customized style based on the mixed latent vector.
8. A system, comprising:
at least one processor; and
at least one memory communicatively coupled to the at least one processor and storing instructions that upon execution by the at least one processor cause the system to perform operations, the operations comprising:
extracting latent vectors from unlabelled data by an encoder, the unlabelled data comprising a plurality of music note sequences, the plurality of music note sequences indicating a plurality of pieces of music;
clustering the latent vectors into a plurality of classes corresponding to a plurality of music styles by a machine learning model;
generating a plurality of labelled latent vectors corresponding to the plurality of music styles by the machine learning model, each of the plurality of labelled latent vectors comprising information indicating features of a corresponding music style among the plurality of music styles; and
generating, by a decoder, a first music note sequence indicating a first piece of music in a particular music style among the plurality of music styles based at least in part on a particular labelled latent vector among the plurality of labelled latent vectors, the particular labelled latent vector corresponding to the particular music style.
9. The system of claim 8 , wherein the encoder comprises an encoder of a variational autoencoder (VAE), wherein the decoder comprises a decoder of the VAE, and wherein the operations further comprising comprise:
inputting the particular labelled latent vector to the decoder of the VAE.
10. The system of claim 9 , wherein the VAE is trained using unlabelled music training data, the encoder of the VAE learns to encode the unlabelled music training data into Gaussian distributions, and the decoder of the VAE learns to decode latent vectors sampled from the Gaussian distributions into pieces of music.
11. The system of claim 8 , wherein the machine learning model comprises a Gaussian Mixture Model (GMM).
12. The system of claim 11 , wherein the GMM is trained using unlabelled latent vectors extracted by an encoder of a VAE from unlabelled music training data.
13. The system of claim 8 , the operations further comprising:
interpolating, by a controllable music generation model, a specific latent vector corresponding to a type of chord sequence with another latent vector corresponding to a random music sequence to generate a mixed latent vector corresponding to a new music sequence, wherein the specific latent vector is among the plurality of labelled latent vectors, and wherein the new music sequence is harmonic with both the chord sequence and the random music sequence.
14. The system of claim 13 , the operations further comprising:
inputting the mixed latent vector to the VAE comprising the decoder; and
generating, by the decoder, a second music note sequence indicating a second piece of music in a customized style based on the mixed latent vector.
15. A non-transitory computer-readable storage medium, storing computer-readable instructions that upon execution by a processor cause the processor to implement operations, the operation comprising:
extracting latent vectors from unlabelled data by an encoder, the unlabelled data comprising a plurality of music note sequences, the plurality of music note sequences indicating a plurality of pieces of music;
clustering the latent vectors into a plurality of classes corresponding to a plurality of music styles by a machine learning model;
generating a plurality of labelled latent vectors corresponding to the plurality of music styles by the machine learning model, each of the plurality of labelled latent vectors comprising information indicating features of a corresponding music style among the plurality of music styles; and
generating, by a decoder, a first music note sequence indicating a first piece of music in a particular music style among the plurality of music styles based at least in part on a particular labelled latent vector among the plurality of labelled latent vectors, the particular labelled latent vector corresponding to the particular music style.
16. The non-transitory computer-readable storage medium of claim 15 , wherein the encoder comprises an encoder of a variational autoencoder (VAE), wherein the decoder comprises a decoder of the VAE, and wherein the operations further comprise:
inputting the particular labelled latent vector to the decoder of the VAE; and
wherein the VAE is trained using unlabelled music training data, the encoder of the VAE learns to encode the unlabelled music training data into Gaussian distributions, and the decoder of the VAE learns to decode latent vectors sampled from the Gaussian distributions into pieces of music.
17. The non-transitory computer-readable storage medium of claim 15 , wherein the machine learning model comprises a Gaussian Mixture Model (GMM).
18. The non-transitory computer-readable storage medium of claim 17 , wherein the GMM is trained using unlabelled latent vectors extracted by an encoder of a VAE from unlabelled music training data.
19. The non-transitory computer-readable storage medium of claim 15 , the operations further comprising:
interpolating, by a controllable music generation model, a specific latent vector corresponding to a type of chord sequence with another latent vector corresponding to a random music sequence to generate a mixed latent vector corresponding to a new music sequence, wherein the specific latent vector is among the plurality of labelled latent vectors, and wherein the new music sequence is harmonic with both the chord sequence and the random music sequence.
20. The non-transitory computer-readable storage medium of claim 19 , the operations further comprising:
inputting the mixed latent vector to a VAE comprising the decoder; and
generating, by the decoder, a second music note sequence indicating a second piece of music in a customized style based on the mixed latent vector.