Source separation using nonnegative matrix factorization with an automatically determined number of bases

1. A method for source separation based on determining a number of bases for a nonnegative matrix factorization (NMF) model, the method comprising:

receiving, at a computing device, a mixed signal including a combination of first signal data and second signal data;

generating, by the computing device, a time-frequency representation of the mixed signal;

determining, by applying a structured stochastic variational inference (SSVI) algorithm to the NMF model, a number of bases for a dictionary of signal-related components of the mixed signal, wherein applying the SSVI algorithm comprises varying the number of bases within a range and selecting the number of bases from the range based on training data;

constructing, using the number of bases and the time-frequency representation, the dictionary and an activation matrix of weights, the weights indicating how active each one of the signal-related components is at a given time; and

separating, using the dictionary and the activation matrix, the first signal data from the second signal data.

2. The method of claim 1 , wherein the determining and the constructing are performed simultaneously.

3. The method of claim 1 , wherein the generating includes creating a power spectrogram based on the first and second signal data, and wherein the signal-related components include vectors of the first signal data and vectors of the second signal data.

4. The method of claim 1 , wherein the first signal data includes a signal of interest emitted by a first source, and wherein the second signal data includes a noise signal emitted by a second source.

5. The method of claim 1 , wherein the separating comprises multiplying the signal-related components and the weights to produce a weighted combination of the signal-related components and weights usable to reconstruct a first signal representing the first signal data. data.

6. The method of claim 1 , further comprising receiving the training data, wherein:

the first signal data includes a first signal from a first text source;

the second signal data includes a second signal from a second text source; and

the training data defines a corpus of words and a plurality of valid sequences of the words of the corpus.

7. The method of claim 1 , further comprising receiving the training data, wherein:

the first signal data includes a first audio signal emitted by a first audio source;

the second signal data includes a second audio signal emitted by a second audio source; and

the training data defines a corpus of sounds and a plurality of valid sequences of the sounds of the corpus.

8. The method of claim 1 , further comprising receiving the training data, wherein the mixed signal is associated with one or more video sources, and wherein the training data defines a corpus of images of interest and a plurality of valid sequences of the images of the corpus.

9. The method of claim 1 , wherein the NMF model is one of a Kullback-Leibler (KL) NMF model, a Bayesian KL-NMF model, and a gamma process KL-NMF model.

10. The method of claim 1 , wherein the NMF model a gamma process NMF (GaPNMF) model.

11. A system for source separation based on determining a number of bases for a nonnegative matrix factorization (NMF) model, the system comprising:

a computing device comprising a processor and a memory having executable instructions stored thereon, that, when executed by the processor, cause the computing device to perform operations comprising:

receiving a mixed signal including a combination of first signal data and second signal data;

generating, by the computing device, a time-frequency representation of the mixed signal;

determining, by applying a structured stochastic variational inference (SSVI) algorithm to the NMF model, a number of bases for a dictionary of signal-related components of the mixed signal, wherein applying the SSVI algorithm comprises varying the number of bases within a range and selecting the number of bases from the range based on training data;

constructing, using the number of bases and the time-frequency representation, the dictionary and an activation matrix of weights, the weights indicating how active each one of the signal-related components is at a given time; and

separating, using the dictionary and the activation matrix, the first signal data from the second signal data.

12. The system of claim 11 , wherein the generating includes creating a power spectrogram based on the first and second signal data, and wherein the signal-related components include vectors of the first signal data and vectors of the second signal data.

13. The system of claim 11 , wherein the first signal data includes a signal of interest emitted by a first source, and wherein the second signal data includes a noise signal emitted by a second source.

14. The system of claim 11 , wherein the separating comprises multiplying the dictionary components and the activation weights to produce a weighted combination of the components and weights usable to reconstruct a first signal representing the first signal data.

15. The system of claim 11 , the operations further comprising receiving the training data, wherein:

the first signal data includes a first signal from a first text source;

the second signal data includes a second signal from a second text source; and

the training data defines a corpus of words and a plurality of valid sequences of the words of the corpus.

16. The system of claim 11 , the operations further comprising receiving the training data, wherein:

the first signal data includes a first audio signal emitted by a first audio source;

the second signal data includes a second audio signal emitted by a second audio source; and

the training data defines a corpus of sounds and a plurality of valid sequences of the sounds of the corpus.

17. The system of claim 11 , the operations further comprising receiving the training data, wherein the mixed signal is associated with one or more video sources, and wherein the training data defines a corpus of images of interest and a plurality of valid sequences of the images of the corpus.

18. The system of claim 11 , wherein the NMF model is one of a Kullback-Leibler (KL) NMF model, a Bayesian KL-NMF model, and a gamma process KL-NMF model, and a gamma process NMF (GaPNMF) model.

19. A non-transitory computer readable storage medium having executable instructions stored thereon, that, when executed by a computing device, cause the computing device to perform operations for source separation based on determining a number of bases for a nonnegative matrix factorization (NMF) model, the instructions comprising:

instructions for receiving a mixed signal including a combination of first signal data and second signal data;

instructions for generating a time-frequency representation of the mixed signal;

instructions for determining, by applying a structured stochastic variational inference (SSVI) algorithm to the NMF model, a number of bases for a dictionary of signal-related components of the mixed signal, wherein applying the SSVI algorithm comprises varying the number of bases within a range and selecting the number of bases from the range based on training data;

instructions for constructing, using the number of bases and the time-frequency representation, the dictionary and an activation matrix of weights, the weights indicating how active each one of the signal-related components is at a given time; and

instructions for separating, using the dictionary and the activation matrix, the first signal data from the second signal data.

20. The non-transitory computer readable storage medium of claim 19 , wherein the dictionary is a matrix of isolated spectra from the mixed signal, and wherein the weights in the activation matrix indicate how active latent components are as a function of time.