Abstract:

Due to the fact that traditional Speech Endpoint Detection methods' performance degrads greatly in a low signal-to-noise ratio and nonstationary noise, a novel robust adatpive threshold endpoint detection method is proposed. First of all, the LSFM parameter is employed as the distinctive feature and the Burg spectrum estimation is applied to figure out the power spectrum, which can enhance the discriminative ability in classifying speech signals and noise, compared with the traditional speech features. Furthermore, an adaptive threshold based on the Bayes estimation criterion is involved in the final judgment, which overcomes the defect of the fixed threshold in adaptability and improves the detection performance to a greater degree. Simulation results show that compared with the traditional feature-based Speech Endpoint Detection methods, the accuracy of the proposed method has a high accuracy rate, which proves that the new method has a better robust performance in a low SNR and nonstationary noise.

Key words: low signal-to-noise ratio, nonstationary noise, speech endpoint detection, long-term spectral flatness measure, Burg spectrum estimation