Endpoint detection algorithm for Mandarin digit recognition using DSP

When Mandarin digit speech recognition (MDSR) is applied in an adverse environment, endpoint detection can be crucial to the entire system. A novel feature-based real-time endpoint detection (FRED) algorithm is proposed. FRED detects the speech endpoint depending on an essential speech feature, which means that the system has high disturbance-proof ability. The feature used for endpoint detection can be figured out conveniently when computing Mel-frequency cepstral coefficients (MFCC); furthermore, the algorithm has low complexity and is suitable for a real-time DSP system. By experiment, the proposed algorithm is shown to be more accurate and noise robust than previously proposed approaches. The FRED algorithm is able to segment vowels and consonants reliably, which makes it possible further to recognize local features to improve system performance.