Method and system of on-the-fly audio source separation

1. A method of audio source separation for an audio mixture, performed by a processor, comprising:

receiving an audio mixture and at least one text query associated with the audio mixture;

retrieving at least one audio sample from an auxiliary audio database by matching the text query with semantic information associated with the auxiliary audio database;

evaluating the at least one audio sample retrieved from the auxiliary audio database; and

separating the audio mixture into a plurality of audio sources using the at least one audio sample.

2. The method of claim 1 , wherein separating the audio mixture uses a subset of the at least one audio sample according to the evaluation result of the audio samples.

3. The method of claim 1 , wherein evaluating the at least one audio sample includes ranking the at least one audio sample according to a match with the received text query.

4. The method of claim 1 , wherein evaluating the audio samples includes ranking the at least one audio sample according to a correlation with the received audio mixture.

5. The method of claim 1 , wherein evaluating the at least one audio sample includes ranking the at least one audio sample according to audio feature similarity between the at least one audio sample and the received audio mixture.

6. The method of claim 1 , wherein the auxiliary audio database includes an internet search engine.

7. The method according to claim 1 , wherein separating the audio mixture includes applying non-negative matrix factorization (NMF) to the audio mixture.

8. The method of claim 1 , wherein evaluating the at least one audio sample and separating the audio mixture are performed jointly.

9. The method of claim 8 , wherein evaluating the at least one audio sample and separating the audio mixture are performed jointly by applying non-negative matrix factorization (NMF) to the audio mixture and the audio samples.

10. The method of claim 9 , wherein evaluating the at least one audio sample is performed by introducing a cost function in the NMF application, wherein the cost function includes a parameter of sparsity penalty so as to allow to choose spectral patterns learned from the retrieved audio examples to guide the decomposition of the mixture.

11. The method of claim 10 , wherein the parameter of sparsity penalty is associated with a trade-off parameter which is adaptively modified in a parameter updating process.

12. A system of audio source separation for an audio mixture, comprising:

a receiving unit configured to receive an audio mixture and at least one text query associated to the audio mixture; and

a processor configured to retrieve at least one audio sample from an auxiliary audio database by matching the text query with semantic information associated to the auxiliary audio database, to evaluate the audio sample retrieved from the auxiliary audio database, and to separate the audio mixture into a plurality of audio sources using the audio sample.

13. The system of claim 12 , wherein the processor separates the audio mixture by applying non-negative matrix factorization (NMF) to the audio mixture.

14. The system of claim 12 , wherein the processor evaluates the audio sample and separates the audio mixture jointly.

15. A non-transitory computer readable medium having stored therein instructions for audio source separation for an audio mixture, which when executed by a computer, cause the computer to:

receive an audio mixture and at least one text query associated with the audio mixture;

retrieve at least one audio sample from an auxiliary audio database by matching the text query with semantic information associated with the auxiliary audio database;

evaluate the at least one audio sample retrieved from the auxiliary audio database; and

separate the audio mixture into a plurality of audio sources using the at least one audio sample.

16. A method of audio source separation for an audio mixture, performed by a processor, comprising:

receiving an audio mixture and at least one text query associated with the audio mixture;

retrieving at least one audio sample from an auxiliary audio database by matching the text query with semantic information associated with the auxiliary audio database;

evaluating the at least one audio sample retrieved from the auxiliary audio database; and

separating the audio mixture into a plurality of audio sources using the at least one audio sample, said evaluating and separating being performed jointly by applying non-negative matrix factorization (NMF) to the audio mixture and the audio samples.