An mcmc approach to joint estimation of clean speech and noise for robust speech recognition

We present a novel framework for joint estimation of speech and noise statistics using a Markov chain Monte Carlo approximation. The underlying distributions of the speech and noise components of noisy speech are estimated at each frame and inferences are made from these distributions. The clean speech is approximated by a discrete distribution, from which new features are extracted and used in the recognition process. The availability of information about the noise statistics enables the algorithm to handle non-stationary noise within an utterance and also improves the overall recognition performance when compared to the previously available sequential Monte Carlo (particle filter) methods for noisy speech compensation. We report experimental results obtained with the Aurora-2 connected digit recognition task and achieve an error reduction of 12.87% over state-of-the-art multi-condition training.