AN ENERGY-BASED ADAPTIVE VOICE DETECTION APPROACH

In this paper we proposed an energy-based adaptive voice detection approach for the front-end of ASR. The main idea of the approach is as follows. Firstly, compute the short-term energy and SNR. Secondly, select the threshold value based on SNR and reference energy based on lower short-term energies. Thirdly, apply the reference energy, the threshold and rules on utterances to locate the voice and the silence in noise. Finally, use smoothing techniques on the detection result to avoid too short silence and voice segments. The proposed approach was used on a small test set of 60 utterances picked from the Switchboard corpus. The baseline approach was the classical energy-based method and the threshold value was in the range of 0.33-0.50. The experimental results showed that the detection accuracy of the proposed approach is about 12% higher than that of the baseline approach in the conditions of SNR>20 dB. The proposed voiced detection approach requires slightly more computation than the classical energy-based approach and can meet the real-time needs