Individualized Denoising of Speech

Speech recorded under normal operating conditions is generally corrupted by varying levels of interfering noise. Noises may arise from different sources – the recording equipment, machines and devices in the recording environment, competing speakers, natural environmental sounds, etc. The noisy speech must ideally be “denoised”, for it to sound better and more intelligible to the listener (or even to an automatic speech recognizer). Consequently, a variety of speech denoising algorithms have been proposed in the literature, to deal with the problem.

The objective of this project is to develop customization algorithms, to customize denoising to individual speakers. We will develop both supervised techniques, where noise-free recordings of the target speaker are available to train the algorithm, and unsupervised methods, where no such prior data are available, and the algorithm must be learned directly from the noisy data. We propose multiple solutions which will be investigated and developed during the proposal period.