Modeling spectral smoothness principle for monaural voiced speech separation

The smoothness of spectral envelope is a commonly known attribute of clean speech. In this study, this principle is modeled through oscillation degree of each time-frequency (T-F) unit, and then incorporated into a computational auditory scene analysis (CASA) system for monaural voiced speech separation. Specifically, oscillation degrees of autocorrelation function (ODACF) and of envelope autocorrelation function (ODEACF) are extracted for each T-F unit, which are then utilized in T-F unit labeling. Experiment results indicate that target units and interference units are distinguished more effectively by incorporating the spectral smoothness principle than by using the harmonic principle alone, and obvious segregation improvements are obtained.