Robust Speaker Recognition Using Approximate Bayesian Inference

This is a demonstration of the application of variational Bayesian inference to the problem of joint speech enhancement and speaker identification. This approach to robust speaker identification illustrates the power of Bayesian inference techniques when applied to signal processing problems of interest. Speech enhancement and speaker identification are important signal processing problems which continue to attract research interest. These problems have traditionally been studied separately but we believe that there are a number of advantages to considering them jointly within a Bayesian framework. A number of authors have presented enhancement algorithms that employ speech and noise priors and use approximate Bayesian techniques for inference. However these prior speech models fail to capture speaker dependent information which could further improve performance at the cost of increased complexity. Also, current approaches to speaker identification mainly rely on directly modeling the speech feature vectors of the speakers to be identified and using clean speech to learn parameters of these models. This approach makes these methods sensitive to noise and as a result these systems do not perform well in real acoustic environments where noise is unavoidable. Furthermore, noise encountered in everyday acoustic environments may include unwanted speech sources with spectral characteristics similar to the desired signal. By considering speech enhancement and identification jointly within a Bayesian framework we can now take advantage of rich speaker dependent speech priors in the enhancement task and appropriately model noise in the identification task.