Maximum likelihood noise HMMm estimation in model-based robust speech recognition

This paper presents a generalization of Rose's Integrated Parametric Model to the gaussian mixture hidden Markov model (HMM), formulation. Observations from clean speech HMM and noise HMM models are combined in the log spectra domain, through a corruption function, to generate noisy speech observations. In order to recognize noisy speech with the proposed model, when only the clean speech HMM and noisy speech adaptation data are available, a maximum likelihood (ML) estimation algorithm for the noise HMM parameters is provided. This algorithm uses the “max” approximation as the corruption function. Noisy digit recognition experiments, with NOISEX-92, show that the same performance is achieved between the proposed model using either a noise model calculated from silent sections of several utterances or the estimated noise model from a single noisy utterance.