Speech Endpoint Detection Based on Improved Cepstral Mean Subtraction

This paper presents a novel endpoint detection method based on Cepstral Mean Subtraction (CMS) for robust and accurate speech recognition in noisy environments. The improved method based on CMS applies Hidden Markov Model (HMM) to do two-step classification for better performance, using the optimal spectral feature subset extracted according to the rule of minimum conditional entropy. In addition, to reduce misrecognition due to the similarity between unvoiced sound and white noise in cepstral feature, we apply weighted smoothing algorithm as a solution. Experiment results show that the proposed method outperforms the conventional approaches in both robustness and accuracy relatively.