Facebook AI Research
Will lecture on
Novel Deep Learning Architectures for Voice Separation and Enhancement
In real-world acoustic environments, a speech signal is frequently corrupted by interfering sounds such as a noisy environment, room conditions, multi-talker setup, etc. The ability to separate and enhance speech forms challenging perceptual tasks and are crucial for any speech processing system designed to perform under such conditions.
In this talk, I'll present some of my recent research around developing and designing novel deep learning architectures for voice separation and speech enhancement. Specifically, I'll present methods to separate an unknown (but bounded) number of speakers, suppress background noise in real-time, and leverage generative models for better speech enhancement. The proposed models work directly over the raw-waveform and significantly outperform the current state-of-the-art methods.