HMM Embedded Conditional Vector Estimation Applied to Noisy Line Spectral Frequencies

Abstract:
Conditional Bayesian estimation is based on disturbed observations, some probability measure of their reliability and a-priori knowledge of the original’s statistics. In this paper an extended conditional vector estimator is developed to account for time-varying statistics of non-stationary vector sources. To this end we introduce an outer HMM, that allows modeling statistics, e.g., phoneme specifically, with particular multivariate Gaussian emissions to provide temporal and spatial correlations for the inner estimator. The scheme is exemplarily applied to enhance noisy line spectral frequencies (LSF). Training and evaluation is done based on an extensive speech database and an AWGN channel model. The results show consistent improvements, with higher gains for low channel SNR.