HMM-supervised classification of the NMF components for robust speech recognition

This paper presents a nonnegative matrix factorization (NMF)-based source separation algorithm for robust speech recognition with music interference. NMF is applied to decompose the mixture signal into a set of basis vectors and corresponding gain vectors, each belonging to either speech or music. Source separation is achieved via classifying the NMF components, i.e. the basis and the corresponding gain vectors into their respective classes. HMM models are incorporated to supervise the classification. More specifically, the likelihood score output from the Viterbi search, i.e. the probability of the input speech given the recognized word models, is adopted as the classification criterion. Such that the separated speech consists of those NMF components having positive contributions to the Viterbi search score. As a result, the recognition output after the separation processing is mostly confident. Automatic speech recognition experiments demonstrate that the proposed source separation algorithm significantly improve the robustness of the recognition system under music interference.