Probabilistic vector mapping of noisy speech parameters for HMM word spotting

A conditional probability model is developed for relating a noisy, observation feature vector to the noise-free vector that generated it. The model is a Gaussian mixture which is based on the vectors and is conditioned on the instantaneous signal-to-noise ratio at the frame. When the feature vector estimates based on this model are used in a hidden Markov model (HMM) word spotter trained with noise-free speech, a performance gain of about 20%-30% is observed (depending on spotter topology) compared to that of the HMM word spotter trained with noisy speech