Adaptive model-based technique for robust speech recognition

This paper presents an algorithm for maximum likelihood (ML) estimation of the noise hidden Markov model (HMM) parameters in the context of model-based techniques. It is applied when only the clean speech HMM and noisy speech adaptation data are available. The noisy speech model is a generalization of Rose\´s (Rose et al. 1994) integrated parametric model to the Gaussian mixture 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. The estimation 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 the estimated noise model from a single noisy utterance or a noise model calculated from silent sections of several utterances.