SNR-Based Inter-Component Phase Estimation Using Bi-Phase Prior Statistics for Single-Channel Speech Enhancement

The fundamental problem of phase-aware single-channel speech enhancement is the estimation of the harmonic phase of signal components from noisy observations. One approach to obtain an estimate of the harmonic phase is by smoothing the noisy harmonic phase in the time domain. Accurate identification of the smoothing regions is crucial to achieve the best enhancement results. Previous works have introduced a binary hypothesis framework for detecting these regions, which is formulated
individually
for each harmonic component based on von Mises statistics and local harmonic SNR levels. Later, smoothing frameworks have been proposed in the
inter-component phase domain
, where the inter-component phases follow the temporal constancy property for clean speech signals. However, these frameworks either do not formulate an SNR-dependent smoothing threshold or impose it empirically. The aim of this work is to formulate an SNR-dependent binary hypothesis framework for smoothing in the
inter-component phase domain
, using prior inter-component phase statistics. The framework then reconstructs the instantaneous phases of the enhanced speech signal components. This proposed framework results in superior speech intelligibility enhancement, reduces phase deviation, and does not introduce auditory buzziness artifacts compared to the previously developed framework that does not formulate an SNR-dependent smoothing threshold.