A phase-dependent a priori SNR estimator in the log-mel spectral domain for speech enhancement

In this work, we address a new single-channel speech enhancement method utilizing phase information. In the conventional speech processing approaches, only the magnitude components were used whereas the phase components were mostly neglected. However, the phase components are known to have some information on speech signals. Hence, we propose a phase-dependent a priori signal-to-noise ratio (SNR) estimator in the log-mel spectral domain to utilize both magnitude and phase information, where the decision-directed (DD) approach is used to determine the a priori SNR from noisy speech. In the computer experiments, the proposed estimator is shown to increase the output SNR of the enhanced speech signals by 2.5 dB compared to the conventional magnitude-based estimator.