Speech Stream Detection for Noisy Environments Based on Empirical Mode Decomposition

A new approach for speech stream detection based on empirical mode decomposition (EMD) under a noisy environment is proposed. Accurate speech stream detection proves to significantly improve speech recognition performance under noise. The proposed algorithm relies on the Teager energy and spectral entropy characteristics of the signal to determine whether an input frame is speech or non-speech. Firstly, the noise signals can be decomposed into different numbers of sub-signals called intrinsic mode functions (IMFs) with the EMD. Then, spectral entropy is used to extract the desired feature for noisy IMF components and Teager energy is used to non-noisy IMF components. Finally, in order to show the effectiveness of the proposed method, we present examples showing that the new measure is more effective than traditional measures. The experiments show that the proposed algorithm can suppress different noise types with different SNR.