A reproducing kernel Hilbert space approach for speech enhancement

The problem of speech enhancement has drawn a considerable amount of research attention over the past few years. Among the techniques developed one finds subspace methods, which seem to offer a good compromise between signal distortion and residual noise level. In this contribution, we present a novel subspace approach to single-channel speech enhancement. The eigen decomposition which was originally performed in the input space is now being done in a reproducing kernel Hilbert space, where the speech nonlinearities can be considered. The proposed algorithm was tested in various noise conditions including white, car, pink and train station noises at various signal-to-noise ratios (SNRs). Objective results show that for white noise, the algorithm presents an average improvement of 73.26% while for colored noise an average improvement of 68.42% is achieved. Subjective tests made on speech, corrupted with white and colored noises, demonstrate that the proposed algorithm provides a significant improvement over other speech enhancement methods found in literature.