In-car speech recognition using model-based wiener filter and multi-condition training

This paper presents in-car speech recognition using a modelbased Wiener filter (MBW) and multi-condition (MC) training. The MBW is a 2-step denoising algorithm based on both rough and precise estimation of speech signals. Correcting roughly estimated signals with a Gaussian mixture model (GMM) makes it possible to accurately denoise with little computational cost. In an evaluation of in-car speech recognition, training of both a GMM and a back-end hidden Markov model (HMM) was performed using both studio-recorded speech signals as well as those signals mixed with in-car noise signals that were recorded in real car environments. In-car speech signals for testing were recorded with a plurality of microphones in different car environments. With respect to word accuracy obtained with MCtrained HMM, it was confirmed that the MBW with MC-trained GMM outperformed the Noise Reduction in ETSI advanced front-end.