Feature Compensation Employing Model Combination for Robust In-Vehicle Speech Recognition

Abstract

An effective feature compensation method is evaluated for reliable speech recognition in real-life in-vehicle environments. The CU-Move corpus, previously collected by RSPG (currently, CRSS) (http://www.utdallas.edu/research/utdrive; Hansen et al., DSP for In-Vehicle and Mobile Systems, 2004), contains a range of speech and noise signals collected for a number of speakers from across the United States under actual driving conditions. PCGMM (parallel combined Gaussian mixture model)-based feature compensation (Kim et al., Eurospeech 2003, 2003; Kim et al., ICASSP 2004, 2004), considered in this chapter, utilizes parallel model combination to generate noise-corrupted speech models by combining clean speech and noise models. In order to address unknown time-varying background noise, an interpolation method of multiple environmental models is employed. To alleviate computational expenses due to multiple models, a noise transition model is proposed, which is motivated from the noise language modeling concept developed in Environmental Sniffing (Akbacak and Hansen, IEEE Trans Audio Speech Lang Process, 15(2): 465–477, 2007). The PCGMM method and the proposed scheme are evaluated on the connected single digits portion of the CU-Move database using the Aurora2 evaluation toolkit. Experimental results indicate that our feature compensation method is effective for improving speech recognition in real-life in-vehicle conditions. Here, a 26.78% computational reduction was obtained by employing the noise transition model with only a slight change in overall recognition performance. The resulting system therefore demonstrates an effective speech recognition strategy for robust speech recognition for noisy in-vehicle environments.