Advanced Feature Normalization and Rapid Model Adaptation for Robust In- Vehicle Speech Recognition

In this study, we present advanced feature normalization and rapid model adaptation for robust in-vehicle speech recognition. For feature normalization, we use a combination of recently established quantile-based cepstral dynamics normalization (QCN) and low pass temporal filtering (RASTALP). Similar to cepstral mean normalization (CMN), QCN aims at alleviating the mismatch between ASR acoustic models and the decoded speech signal. QCN relaxes CMN assumptions concerning feature distributions, making the normalization more stable in varying adverse environments. RASTALP is a low-pass approximation of RASTA filtering which significantly reduces transient distortions introduced by the original band-pass filter. Using the normalized features, we adapt the speaker-independent acoustic model to specific speakers. The adaptation method is based on an aspect model (a “mixture-ofmixtures” model). To enable adaptation requiring only extremely small amounts of adaptation data (i.e., a few seconds), we train a small number of mixture models which can be interpreted as models of probabilistic “speaker clusters” for in-vehicle environments. In this work, we use fMLLR to represent individual speaker models. The speaker models are mixed using weights determined from adaptation data. Experimental results show that the normalization employing QCN-RASTALP is consistently superior to CMN. We also observe that in contrast to the conventional methods, the adaptation based on the aspect model improves word error rates for the in-vehicle noise environments.