Speech Activity Detection Method Based on Spectral Information with Multi-Scale Frame Length

Speech activity detection (SAD) is a fundamental study that has been widely applied to the preprocessing steps of various speech algorithms. As influenced by the choice of frame length, the detection effect of nondata-driven SAD methods under low signal-to-noise ratios (SNR) frequently does not present well the ability of energy, entropy, and variance, which are spectral features. In this paper, we propose to combine the spectral information at multiscale frame length conditions with the matrix dimensionality reduction method to enhance the discrimination between noise and speech at low SNR while providing a new perspective for frame length selection in speech processing. The experimental results show that the proposed method based on the combination of spectral energy, entropy, and variance with multi-scale frame length outperforms the energy-to-entropy ratio and the variance-to-entropy ratio methods at a certain single scale under SNR values from -10 dB to 10 dB.