Speech Dereverberation using EM Algorithm and Kalman Filtering

Reverberation is a typical acoustic phenomenon in enclosures, that usually deteriorates the speech signal due to the multitude of reflections from the walls, ceiling, and large objects. In general, dereverberation algorithms aim at the reduction of these reflections, which emphasizes the original speech signal and improves both its quality and intelligibility.

For the systematic development of dereverberation algorithms, we define a statistical model for the speech signal, and for the acoustic room impulse response. The expectation-maximization (EM) approach is then applied to this model, where the clean speech is estimated in the E-step using a Kalman smoothing, and the acoustic parameters are updated in the M-step. For online applications and dynamic scenarios, i.e., where the speaker and/or the microphones are moving, we derived a recursive-EM (REM) algorithm which uses an online Kalman filter instead of a Kalman smoother, and which updates the parameters by using only the new observed data. 

Two extensions of this approach were developed as well. The first extension is a segmental algorithm, where iterations are performed over short segments of the recording. This way, the latency of the algorithm is controlled, while the accuracy can be iteratively improved. The second extension is a binaural algorithm mainly applicable to hearing aids. The binaural algorithm trades off between the reduction of reverberation and the preservation of the spatial perception of the user.

* Ph.D. research supervised by Prof. Sharon Gannot and Prof. Emanuël Habets