Acoustic source separation systems

1. A method for acoustic source separation, the method comprising:

inputting acoustic data from a plurality of acoustic sensors, said acoustic data comprising acoustic signals combined from a plurality of acoustic sources;

converting said acoustic data to time-frequency domain data comprising a plurality of time-frequency data frames for a plurality of times and frequencies; and

constructing a multichannel filter to operate on said time-frequency data frames to separate signals from said acoustic sources;

wherein said constructing comprises:

determining a set of de-mixing matrices (W f ), one for each of said plurality of frequencies, to apply to each said time-frequency data frame to determine a vector of separated outputs ( y ft ), wherein said determining comprises modifying each of said de-mixing matrices by a respective gradient value (G;G′) for a said frequency dependent upon a gradient of a cost function measuring a separation of said sources by the respective said de-mixing matrix; and

wherein said respective gradient values for each frequency are each calculated from a stochastic selection of said time-frequency data frames,

wherein said stochastic selection of said time-frequency data frames comprises randomly selecting a plurality, L, of said time-frequency data frames for determining each of said de-mixing matrices (W f ).

2. A method as claimed in claim 1 wherein the same value L is used for determining a plurality or all of said de-mixing matrices (W f ) for said plurality of frequencies.

3. A method as claimed in claim 1 further comprising selecting said time-frequency data frames according to a selection probability λ t associated with each frame, where λ t represents the probability of selecting a time-frequency data frame for time t.

4. A method as claimed in claim 3 further comprising determining a value for λ t dependent upon an importance metric for the time-frequency data frame for time t.

5. A method as claimed in claim 4 further comprising determining source activity data from said acoustic data, and determining said importance metric λ t for a time-frequency data frame from said source activity data.

6. A method as claimed in claim 5 wherein said importance metric is defined to equalise contributions from sources with different numbers of frames in which the sources are active.

7. A method as claimed in claim 1 comprising iteratively modifying each of said de-mixing matrices by said respective gradient value, and varying the number L of stochastically selected time-frequency data frames from one iteration to another.

8. A method as claimed in claim 1 further comprising prescaling said set of de-mixing matrices (W f ) to minimise a substitute cost function prior to constructing said multichannel filter.

9. A method of blind source separation using the method of claim 1 to construct said multichannel filter;

applying said multichannel filter to said time-frequency domain data to determine de-mixed time-frequency data representing said acoustic sources; and

converting said de-mixed time-frequency data to the time domain to recover de-mixed time domain data for at least one of said acoustic sources.

10. A non-transitory storage medium storing processor control code to implement the method of claim 1 .

11. A system for acoustic source separation, the system comprising:

one or more inputs to receive acoustic data from a plurality of acoustic sensors, said acoustic data comprising acoustic signals combined from a plurality of acoustic sources; and

an audio signal processor coupled to said one or more inputs, the audio signal processor comprising:

a time-to-frequency domain converter to convert said acoustic data to time-frequency domain data comprising a plurality of time-frequency data frames for a plurality of times and frequencies; and

a multichannel filter module to operate on said time-frequency data frames to separate signals from said acoustic sources; and

wherein said audio signal processor is further configured to:

determine a set of de-mixing matrices (W f ), one for each of said plurality of frequencies, to apply to each said time-frequency data frame to determine a vector of separated outputs ( y ft ), by modifying each of said de-mixing matrices by a respective gradient value (G;G′) for a said frequency dependent upon a gradient of a cost function measuring a separation of said sources by the respective said de-mixing matrix; and to

calculate said respective gradient values for each frequency from a stochastic selection of said time-frequency data frames,

wherein said stochastic selection of said time-frequency data frames comprises randomly selecting a plurality, L, of said time-frequency data frames for determining each of said de-mixing matrices (W f ).