Speeding Up Permutation Invariant Training for Source Separation
This work addresses a computational bottleneck for researchers and practitioners in audio source separation, making PIT more scalable, though it is incremental as it builds on existing PIT methods.
The paper tackles the exponential complexity of Permutation Invariant Training (PIT) in source separation, which limits its use for many speakers or long recordings, by decomposing the PIT criterion to enable polynomial-time algorithms like the Hungarian algorithm, reducing computational costs.
Permutation invariant training (PIT) is a widely used training criterion for neural network-based source separation, used for both utterance-level separation with utterance-level PIT (uPIT) and separation of long recordings with the recently proposed Graph-PIT. When implemented naively, both suffer from an exponential complexity in the number of utterances to separate, rendering them unusable for large numbers of speakers or long realistic recordings. We present a decomposition of the PIT criterion into the computation of a matrix and a strictly monotonously increasing function so that the permutation or assignment problem can be solved efficiently with several search algorithms. The Hungarian algorithm can be used for uPIT and we introduce various algorithms for the Graph-PIT assignment problem to reduce the complexity to be polynomial in the number of utterances.