Surrogate Source Model Learning for Determined Source Separation
Please refer to its documentation for more details.
Abstract
We propose to learn surrogate functions of universal speech pri-ors for determined blind speech separation. Deep speech priorsare highly desirable due to their superior modelling power, but arenot compatible with state-of-the-art independent vector analysisbased on majorization-minimization (AuxIVA), since deriving therequired surrogate function is not easy, nor always possible. In-stead, we do away with exact majorization and directly approximatethe surrogate. Taking advantage of iterative source steering (ISS)updates, we back propagate the permutation invariant separationloss through multiple iterations of AuxIVA. ISS lends itself well tothis task due to its lower complexity and lack of matrix inversion.Experiments show large improvements in terms of scale invariantsignal-to-distortion (SDR) ratio and word error rate compared tobaseline methods. Training is done on two speakers mixtures andwe experiment with two losses, SDR and coherence. We find thatthe learnt approximate surrogate generalizes well on mixtures ofthree and four speakers without any modification. We also demon-strate generalization to a different variation of the AuxIVA updateequations. The SDR loss leads to fastest convergence in iterations,while coherence leads to the lowest word error rate (WER). Weobtain as much as36 %reduction in WER.
Authors
- Robin Scheibler
- Masahito Togami