Convolutive Prediction for Monaural Speech Dereverberation and Noisy-Reverberant Speaker Separation

A promising approach for speech dereverberation is based on supervised learning, where a deep neural network (DNN) trained to predict the direct sound from noisy-reverberant speech. This data-driven leveraging prior knowledge of clean patterns, and seldom explicitly exploits linear-filter structure in reverberation, i.e., that reverberation results linear convolution between room impulse response (RIR) dry source signal. In this work, we propose exploit within learning monaural framework. The key idea first estimate direct-path signal target speaker using DNN then identify signals are decayed delayed copies estimated signal, as these can be reliably considered reverberation. They either directly removed dereverberation, or used extra features another perform better dereverberation. To copies, underlying filter (or RIR) by efficiently solving regression problem per frequency time-frequency domain. We modify proposed algorithm separation reverberant conditions. State-of-the-art obtained REVERB, SMS-WSJ, WHAMR! datasets.