Robust Automatic Speech Recognition through On-line Semi Blind Signal Extraction

This paper describes the system used to process the data of the CHiME Pascal 2011 competition, whose goal is to separate the desired speech and recognize the commands being spoken. The binaural recorded mixtures are processed by an on-line Semi- Blind Source Extraction algorithm. The algorithm is based on a multi-stage architecture combining the advantages of con- strained Independent Component Analysis and Wiener-based processing, allowing the estimation of the target signal with lim- ited distortion. The recovered target signal is then fed to the rec- ognizer which uses noise robust features based on Gammatone Frequency Cepstral Coefficients. Moreover, model adaptation to actual processing is applied as a further stage to reduce the acoustic mismatch. Performance comparison between differ- ent model/algorithmic settings is reported for both development and test data sets.