NOISE ROBUST DISTANT AUTOMATIC SPEECH RECOGNITION UTILIZING NMF BASED SOURCE SEPARATION AND AUDITORY FEATURE EXTRACTION
Niko Moritz, Marc René Schädler, Kamil Adiloğlu, Bernd T. Meyer, Tim Patrick Jürgens, Timo Gerkmann, Birger Kollmeier, Simon Doclo, Stefan Goetze · 2013
This paper describes our contribution to the 2 nd CHiME challenge and focuses on the small vocabulary task, i.e. track one. We present a robust system combination that involves source separation, auditory feature extraction and a modified automatic speech recognition back-end. The source separation code is based on a non-negative matrix factorization approach and the presented auditory feature extraction method uses 2D Gabor filter functions to extract spectral, temporal and spectro-temporal information of the speech signals. In addition we describe the modifications to our classification back-end and discuss the achieved results. On the final CHiME test set the proposed system achieves a maximum keyword recognition rate improvement of 50.25 % for the-6 dB SNR condition, for instance. Index Terms — CHiME challenge, non-negative matrix factorization, Gabor feature extraction, source separation, automatic speech recognition 1.