Robust speech recognition using model-based feature enhancement and harmonic decomposition in the missing data framework

Missing Data Theory (MDT) has shown to improve the robustness of automatic speech recognition (ASR) systems. A crucial part in a MDT-based recogniser is the computation of the reliability masks from noisy data. For each component of the feature vector extracted from the noisy observation, the Missing Data Detector (MDD) makes a decision about the presence of speech and noise. The components that are likely to be dominated by the noise, are labelled as unreliable and their values will be estimated from the reliable observations. In this paper, we exploit the Model-Based Feature Enhancement (MBFE) technique in the reliability decisions of the MDD. This technique makes use of statistical models of clean speech and noise to generate estimates of the original speech and noise feature vectors. Decision criteria obtained by combining the MBFE-estimates with the components derived from the harmonic decomposition of the noisy speech signal showed a further increase in performance of the recognition system operating on noisy signals.status: publishe