Speech Recognition Using Factorial Hidden Markov Models for Separation in the Feature Space

The following audio examples demonstrate the performance of the algorithm submitted to the speech separation challenge and ICSLP 2006 conference. Each mixture sample was randomily selected from the database and separated into target and masker signals.

mixture signal
(signal-to-masker level)
separated target separated masker
orig1.wav (6 dB) sep1_1.wav sep1_2.wav
orig2.wav (3 dB) sep2_1.wav sep2_2.wav
orig3.wav (0 dB) sep3_1.wav sep3_2.wav
orig4.wav (-3 dB) sep4_1.wav sep4_2.wav
orig5.wav (-6 dB) sep5_1.wav sep5_2.wav
orig6.wav (-9 dB) sep6_1.wav sep6_2.wav

Demonstrations main page

- Tuomas Virtanen, tuomas.virtanen@tut.fi