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

摘要

Missing Data Theory (MDT) has shown to improve the robustnessof automatic speech recognition (ASR) systems. A crucialpart in a MDT-based recogniser is the computation of thereliability masks from noisy data. For each component of thefeature vector extracted from the noisy observation, the MissingData Detector (MDD) makes a decision about the presence ofspeech and noise. The components that are likely to be dominatedby the noise, are labelled as unreliable and their values willbe estimated from the reliable observations. In this paper, we exploitthe Model-Based Feature Enhancement (MBFE) techniquein the reliability decisions of the MDD. This technique makesuse of statistical models of clean speech and noise to generateestimates of the original speech and noise feature vectors. Decisioncriteria obtained by combining the MBFE-estimates withthe components derived from the harmonic decomposition of thenoisy speech signal showed a further increase in performance ofthe recognition system operating on noisy signals.