STFT Phase Improvement for Single Channel Speech Enhancement

Abstract:
In state-of-the-art single channel short-time Fourier transform (STFT) based speech enhancement algorithms only the amplitude of the noisy speech signal is improved, but its phase is left unchanged. It is commonly assumed that the noisy phase is the best estimate of the clean phase available. While using the noisy phase is indeed optimal under certain statistical assumptions, in this paper we show that blindly improving the noisy phase is possible when these, potentially limiting, assumptions are dropped. Without modifying the amplitude, the proposed algorithm leads to frequency weighted SNR improvements of up to 1.8 dB. Further, the presented phase enhancement scheme is real-time capable and can be combined with any off-the-shelf STFT-based amplitude estimator. Index Terms — speech enhancement, phase estimation, noise reduction, signal reconstruction