LINEAR PREDICTION BASED DEREVERBERATION FOR SPHERICAL MICROPHONE ARRAYS

Dereverberation is an important preprocessing step in many speech systems, both for human and machine listening. In many situations, including robot audition, the sound sources of interest can be incident from any direction. In such circumstances, a spherical microphone array allows direction of arrival estimation which is free of spatial aliasing and direction-independent beam patterns can be formed. This contribution formulates the Weighted Prediction Error algorithm in the spherical harmonic domain and compares the performance to a space domain implementation. Simulation results demonstrate that performing dereverberation in the spherical harmonic domain allows many more microphones to be used without increasing the computational cost. The benefit of using many microphones is particularly apparent at low signal to noise ratios, where for the conditions tested up to 71% improvement in speech-to-reverberation modulation ratio was achieved.