Mixtures of Gamma Priors for Non-Negative Matrix Factorization Based Speech Separation

These audio examples demonstrate the performance of the algorithm to be presented at the ICA 2009 conference. Here, a model for male speakers and a model for female speakers was trained without using the test speakers. The mixtures were separated using the proposed method with 30 spectrum vectors, with and without adaptation of the vectors. The signals were generated by mixing random sentences from the GRID corpus.

mixture signal
separated female (fixed basis vectors) separated male (fixed basis vectors) separated female (adaptive bases) separated male (adaptive bases)
mix1.wav 30sepfixed1-1.wav 30sepfixed1-2.wav 30sepadaptive1-1.wav 30sepadaptive1-2.wav
mix2.wav 30sepfixed2-1.wav 30sepfixed2-2.wav 30sepadaptive2-1.wav 30sepadaptive2-2.wav
mix3.wav 30sepfixed3-1.wav 30sepfixed3-2.wav 30sepadaptive3-1.wav 30sepadaptive3-2.wav

The following signals were obtained by using 70 basis vectors:
mixture signal
separated female (fixed basis vectors) separated male (fixed basis vectors) separated female (adaptive bases) separated male (adaptive bases)
mix1.wav 70sepfixed1-1.wav 70sepfixed1-2.wav 70sepadaptive1-1.wav 70sepadaptive1-2.wav
mix2.wav 70sepfixed2-1.wav 70sepfixed2-2.wav 70sepadaptive2-1.wav 70sepadaptive2-2.wav
mix3.wav 70sepfixed3-1.wav 70sepfixed3-2.wav 70sepadaptive3-1.wav 70sepadaptive3-2.wav

Demonstrations main page

- Tuomas Virtanen, tuomas.virtanen@tut.fi