HMM modeling of additive noise in the western languages context

This paper is concerned to the noisy speech HMM modelling when the noise is additive, speech independent and the spectral analysis is based on sub-bands. The internal distributions of the noisy speech HMM’s were derived when Gaussian mixture density distributions for clean speech HMM modelling are used, and the noise is normally distributed and additive in the time domain. In these circumstances it is showed that the HMM noisy speech distributions are not Gaussians, however, fitting these distributions as a Gaussian mixture, only a little bit of loss in performance was obtained at very low signal to noise ratios, when compared with the case where the real distributions were computed using Monte Carlo methods.