Joint speech enhancement and speaker identification using approximate Bayesian inference

A variational Bayesian principle is applied to derive a iterative technique for jointly identifying a speaker in a noisy acoustic environment and enhancing their speech. Intuitively, it is clear that employing speaker dependent priors for speech allows for better speech enhancement, while cleaner speech allows for better speaker identification. The derived algorithm reflects this intuition by iteratively exchanging information between the enhancement and identification tasks. Experimental results using the TIMIT data set are presented to demonstrate the algorithm's performance.