On the importance of harmonic phase modification for improved speech signal reconstruction

Abstract

Conventional single-channel speech enhancement is mainly focused on modifying the noisy short-time Fourier transform amplitude spectrum while for signal reconstruction the noisy phase is used. Recent advances demonstrate the positive improvements in speech enhancement when the noisy phase is replaced with an estimated clean phase for signal reconstruction. In this paper, we study the impact of the linear phase and unwrapped phase components provided by harmonic phase decomposition on the speech quality at signal reconstruction. We present objective and subjective results comparing the proposed harmonic phase modification with other phase estimation methods. Our results show that enhancement of decomposed phase parts suffices for improved reconstruction in speech enhancement.