Missing Feature reconstruction with Multivariate Laplace distribution (MLD) for noise robust phoneme recognition

Speech recognition accuracy degrades in presence of additive noise especially when recognizer's training data are clean. Several methods have been proposed to compensate effects of noise on recognition accuracy among these methods missing feature techniques (MFT) have shown promising results in making speech recognition systems robust. In MFT spectral vectors of speech signal are conventionally modeled with a Gaussian distribution (GD). In this paper we proposed to model distribution of spectral vectors of speech signals with a multivariate Laplace distribution (MLD) in order to reconstruct missing elements. First we mathematically estimate missing values of each spectral vector conditioned on observed values with MLD parameters and by using this estimate in reconstruction of missing features we improved phoneme recognition accuracy up to 13.62% in -lOdB and 11.1% in 0 dB in presence of different additive noises in comparison with modeling spectral vectors with a Gaussian distribution (GD).