Robust endpoint detection for in-car speech recognition

The endpoint detection plays a significantly important role in the front end processing of speech recognition. It is very difficult, however, to precisely locate endpoints on the input utterance to be free on non-speech regions because of unpredictable background noise. This paper proposes a novel approach that finds robust features for better endpoint detection in a noisy incar environment. In the proposed method, we integrate both the widely used energy and entropy [2], [6] to form a new feature that possesses advantages of each individual while compensating the drawback of each other. By monitoring the variation of the extracted new features, more precise endpoints can be found. Experimental results present that this algorithm outperforms the energy-based algorithms in both accuracy of boundary point detection and recognition performance under a real in-car noisy environment. The result of accuracy improvement shows 10% higher comparing with energy-based algorithm.