Noise adaptation of HMM speech recognition systems using tied-mixtures in the spectral domain

We compare two different approaches to the problem of additive noise in a hidden Markov model (HMM) filterbank-based speech recognition system: (i) preprocessing by estimation and (ii) adaptation of the HMM output probability distributions. The adaptation method, previously formulated only for the static spectral features, is generalized in this paper to the time-derivative of the spectrum. Estimation and adaptation are formulated with a common statistical model (MIXMAX) and are compared using the same recognition system. We find that under low and medium signal-to-noise ratio (SNR) conditions, parameter adaptation is superior to preprocessing by estimation