Evaluation of Robust Speech Recognitio Speech Recognition in a Noisy Aut

In this paper, we evaluate the performance of several robust speech recognition algorithms in a noisy automobile environment as characterized by the Finnish SpeechDat–Car ASR task [1]. By applying acoustic feature compensation, model compensation, and speech detection algorithms to this task, a 51% reduction in word error rate (WER) was obtained relative to the ETSI standard ASR front–end. In addition, these same techniques achieved an average 35% WER reduction for clean condition training and multiple condition training on a simulated speech–in–noise task as characterized by the Aurora 2 ASR task [2]. The paper also presents alternatives for how these algorithms can be implemented in a distributed speech recognition framework.