Study On Endpoint Detection Methods Of Speeh Signal

Two kind of the methods about endpoint detection: an algorithm based on the spectrum entropy method in endpoint detection and an algorithm based on the double threshold comparing method in endpoint detection in this paper. Voice activity detection in low signal-to-noise ratio environments is improved with an algorithm based on he spectrum entropy. Each frame is first divided into 16 bands with selection of bands with frequencies between 250Hz and 4kHz and energies below 90% of the total energy. The energy and the SNR of each band after speech enhancement are then calculated with the entropy band weight adjusted according to it's SNR. The smoothed entropy is then used for the voice activity detection. Test results show that the method significantly increases the voice activity detection ratio. Based on study of the double threshold comparing method in endpoint detection, the paper proposed the three-time searching endpoint detection of speak signals with the threshold unchanged. The method include three steps: multi-words detection, endpoints-restore and leak-points detection. The problem of threshold setting exists in the double-threshold comparing detection in continuous speech is solved effectively in this method, with the normalized speech signals, it can detect speech endpoints accurately at the same threshold. The endpoint detection based on short-time average magnitude of speech signals' relative autocorrelation sequences can detect in high accuracy under noise conditions.