Modelling, from physics to signal processing Segmentation and separation of speech and noise signals using coherence and power spectrum functions

摘要

The available observed signals are convolutive mixtures of various source signals. The sources are extracted by recombining the observed signals using ratios of mixing filters, identified during silence phases of one source. The identification of these ratios of mixing filters uses estimates of the PSD and CPSD of the observations provided by a modified version of averaged periodogram. Silence phases are detected using the mean of the segmented real coherence function of the observations over the frequency range [0-800 Hz].