Non-stationary Environment Compensation Using Sequential EM Algorithm for Robust Speech Recognition

Abstract

The paper presents a non-stationary environment compensation using sequential EM estimation for tracking the complicated environment. All of the noisy features used in the recognition system are effectively compensated. The speech corruption in the log domain such as the 24 log-filterbank coefficients and the log-energy feature can be modeled as a nonlinear model. For efficient estimating noise parameter using the subsequent sequential Expectation-Maximization (EM) algorithm, the nonlinear environment model is linearized by the truncated first-order vector Taylor series (VTS) approximation. Due to the cepstral features are nearly independence, we train the clean speech using cepstral features and the log-energy feature, and then obtain a diagonal Gaussian mixture model in the log domain by taking inverse discrete cosine transform (IDCT). The experiments are conducted on the large vocabulary continuous speech recognition (LVCSR) system. Results demonstrate that it achieves attractive improvements when compared with CMN (cepstral mean normalization) and the batch-EM based compensation approach.