Speech Endpoint Detection Based on the Dynamic Segmentation of Power Spectral Envelope

The acoustic feature is not robust enough due to the interference of environmental noises. Some heuristic approaches of smoothing noisy spectra were introduced to treat with this problem. But those methods did not consider the intrinsic correlation in the time domain. This paper presents a novel method of endpoint detection, where the time sequence of logarithmic power was partitioned into homogeneous blocks using dynamic auto-segmentation. The acoustic feature was extracted from each homogenous block. The endpoint detection was conducted based on the unit of homogenous block. The experimental results showed the superiority of the proposed method.