Preprocessing of Independent Vector Analysis Using Feed-Forward Network for Robust Speech Recognition

摘要

This paper describes an algorithm to preprocess independent vector analysis (IVA) using feed-forward network for robust speech recognition. In the framework of IVA, a feed-forward network is able to be used as an separating system to accomplish successful separation of highly reverberated mixtures. For robust speech recognition, we make use of the cluster-based missing feature reconstruction based on log-spectral features of separated speech in the process of extracting mel-frequency cepstral coefficients. The algorithm identifies corrupted time-frequency segments with low signal-to-noise ratios calculated from the log-spectral features of the separated speech and observed noisy speech. The corrupted segments are filled by employing bounded estimation based on the possibly reliable log-spectral features and on the knowledge of the pre-trained log-spectral feature clusters. Experimental results demonstrate that the proposed method enhances recognition performance in noisy environments significantly.