Research intern talk: Unified speech enhancement approach for speech degradation & noise suppression

Опубликовано 11 ноября 2023, 0:50

Speakers: Khandokar Md. Nayem
Host: Sebastian Braun

Speech enhancement approaches generally focus on removing additive noise and reverberation that adversely affects the overall speech quality and intelligibility. Another group of signal degradations like clipping, bandwidth limitations, and codec degradation can occur due to poor recording hardware, network transmission, and other pre-processing. These degradations largely impact on intelligibility and speech quality. In this work, we deploy a convolutional recurrent network to remove these speech degradations in conjunction with the noise suppression task and propose cascade and end-to-end approaches. We compare both complex mask and direct spectrum estimation approaches for this task using a small real-time capable DNN. Overall, we propose a cascaded processing approach, addressing the distortion types differently, and enabling a task-tailored modular processing.

Learn more: microsoft.com/en-us/research/v...