Dereverberation of autoregressive envelopes for far-field speech recognition

The task of speech recognition in far-field environments is adversely affected by the reverberant artifacts that elicit as temporal smearing sub-band envelopes. In this paper, we develop a neural model for dereverberation using long-term envelopes speech. are derived frequency domain linear prediction (FDLP) which performs an autoregressive estimation Hilbert estimates envelope gain when applied to signals suppresses late reflection components signal. dereverberated used feature extraction recognition. Further, sequence steps involved dereverberation, and acoustic modeling ASR can be implemented single processing pipeline allows joint learning network model. Several experiments performed on REVERB challenge dataset, CHiME-3 dataset VOiCES dataset. these experiments, yields significant performance improvements over baseline system based log-mel spectrogram well other past approaches (average relative 10–24% system). A detailed analysis choice hyper-parameters cost function also provided.