Robust speech/non-speech detection in adverse conditions based on noise and speech statistics

Recognition performance decreases when recognition systems are used over the telephone network, especially wireless network and noisy environments. It appears that non efficient speech/non-speech detection is a very important source of this degradation. Therefore, speech detector robustness to noise is a challenging problem to be examined, in order to improve recognition performance for the very noisy communications. Speech collected in GSM environment gives an example of such very noisy speech to be recognized. Several studies were conducted aiming to improve the robustness of speech/non-speech detection used for speech recognition in adverse conditions. This paper introduces a robust word boundary detection algorithm reliable in the very noisy cellular network environment. The algorithm is based on the statistics of noise and speech in the observed signal. In order to decide on the binary hypotheses of noise only versus speech plus noise, we use a likelihood ratio criterion.