Likelihood-Maximizing-Based Multiband Spectral Subtractionfor Robust Speech Recognition

摘要

Automatic speech recognition performance degrades significantly when speech is affected by environmental noise. Nowadays, themajor challenge is to achieve good robustness in adverse noisy conditions so that automatic speech recognizers can be used in realsituations. Spectral subtraction (SS) is a well-known and effective approach; it was originally designed for improving the qualityof speech signal judged by human listeners. SS techniques usually improve the quality and intelligibility of speech signal whilespeech recognition systems need compensation techniques to reduce mismatch between noisy speech features and clean trainedacoustic model. Nevertheless, correlation can be expected between speech quality improvement and the increase in recognitionaccuracy. This paper proposes a novel approach for solving this problem by considering SS and the speech recognizer not as twoindependent entities cascaded together, but rather as two interconnected components of a single system, sharing the common goalof improved speech recognition accuracy. This will incorporate important information of the statistical models of the recognitionengine as a feedback for tuning SS parameters. By using this architecture, we overcome the drawbacks of previously proposedmethods and achieve better recognition accuracy. Experimental evaluations show that the proposed method can achieve significantimprovement of recognition rates across a wide range of signal to noise ratios.