Permutation invariant training for talker-independent multi-talker speech separation

A system for separating two or more audio source signals from a first monaural signal having audio source signals and noise source signals, the system comprising:

• one or more processors;
• and
• a memory having stored thereon computer-executable instructions that, when executed by the one or more processors, configure the processors to:
• generate, from a window of frames of a second monaural signal, output layers comprising estimates of audio source signals attributable to disparate audio signal sources contributing to the second monaural signal;
• generate a plurality of labels, wherein a total number of the plurality of labels is equal to a total number of the disparate audio signal sources;
• iteratively assign the plurality of labels to the disparate audio signal sources for all possible combinations of labels and disparate audio signal sources to create a set of possible assignments, each possible assignment in the set of possible assignments corresponding to a combination of labels and disparate audio signal sources;
• obtain a plurality of spatially filtered signals, wherein a total number of spatially filtered signals equal to the total number of the plurality of labels;
• determine assignment error scores for each of the set of possible assignments, the assignment error scores determined based at least in part on a difference between labels of the plurality of labels for respective output layers for a respective possible assignment and the plurality of spatially filtered signals;
• determine which individual audio signal sources to associate with individual estimates based at least in part on a minimum total deviation between training data associated with the individual audio signal sources and the individual estimates, wherein the minimum total deviation is a lowest assignment error of the assignment error scores;
• generate a set of masks by iteratively modifying parameters of the model to reduce the minimum total deviation;
• and
• generate the two or more audio source signals from the first monaural signal by using the set of masks, the source of the two or more audio source signals being different from a source of the audio source signals in the second monaural signal.

A method for separating two or more audio source signals from a first monaural signal having audio source signals and noise source signals, the method comprising:

• generate, from a window of frames of a second monaural signal, output layers comprising estimates of audio source signals attributable to disparate audio signal sources contributing to the second monaural signal;
• generating a plurality of labels, wherein a total number of the plurality of labels is equal to a total number of the disparate audio signal sources;
• iteratively assigning the plurality of labels to the disparate audio signal sources for possible combinations of labels and disparate audio signal sources to create a set of possible assignments, each possible assignment in the set of possible assignments corresponding to a combination of labels and disparate audio signal sources;
• obtaining a plurality of spatially filtered signals, wherein a total number of spatially filtered signals equal to the total number of the plurality of labels;
• determining assignment error scores for each of the set of possible assignments, the assignment error scores determined based at least in part on a difference between labels of the plurality of labels for respective output layers for a respective possible assignment and the plurality of spatially filtered signals;
• determining an assignment pattern for the estimates, the assignment pattern defining which individual audio signal sources are associated with which individual estimates based at least in part on a minimum total pairwise deviation between training data associated with the individual audio signal sources and the individual estimates, wherein the minimum total pairwise deviation is a lowest assignment error score of the assignment error scores;
• generate a set of masks by iteratively modifying parameters of the model to reduce the minimum total deviation;
• and
• generate the two or more audio source signals from the first monaural signal by using the set of masks and the assignment pattern, the source of the two or more audio source signals being different from a source of the audio source signals in the second monaural signal.