A wavelet denoising technique to improve endpoint detection in adverse conditions

Recognition performance decreases when automatic recognition systems are used over the telephone network, especially wireless network and noisy environments. Previous studies have shown that non efficient speech/non-speech detection is a very important source of this degradation. Hence, speech detector robustness to noise is highly required, in order to improve recognition performance for the very noisy communications. Several studies were conducted aiming at increasing the robustness of speech/non-speech detection used for speech recognition in adverse conditions. However, despite the improvements, many segments of noise may be wrongly detected by the robust speech/non-speech detector, which increases the false acceptance errors. Therefore, this paper introduces an efficient method to reject such false detections in order to provide a robust word boundary detection algorithm reliable in the very noisy cellular network environment. The algorithm is based on a denoising technique using a discrete wavelet transform of the detector's output segments.