On the importance of phase and magnitude spectra in speech enhancement

The aim of this paper is to investigate the importance of phase and magnitude spectra in speech enhancement at different conditions with emphasizing on the role of phase spectrum. The speech signal is exposed to additive noise in different SNRs. Then, it is decomposed into different frame lengths from 32 to 1024 ms. In synthesis stage we have used clean phase spectrum along with noisy magnitude spectrum as well as clean magnitude spectrum along with noisy phase spectrum. The quality of speech is evaluated by PESQ objective measure. The maximum speech quality improvement in SNR of 0 dB in case of using clean phase is 1.1 in PESQ scale obtained in frame length of 128 ms and in case of using clean magnitude spectrum is 2.2 in PESQ scale in frame length of 32 ms. Finally, we have shown that phase spectrum cleaning for female speakers leads to more speech quality improvement and in case of male speakers magnitude spectrum enhancement is more useful.