The long-term information of speech signals shows excellent performances in the applications of voice activity detection. Six types of long-term information based on auditory filter banks are proposed through the non-linear spectral decomposition with three different auditory filters. Further

an adaptive voice activity detection algorithm based on these types of long-term information is proposed. Without additional training data

this algorithm use the data selecting from the test signals according to long-term information to train a speech/non-speech classifier

and classifies the current test signals using the speech/non-speech classifier frame by frame. Experiments on TIMIT dataset and NOISEX-92 dataset show that the algorithm improves the performance of VAD with higher accuracy and stronger robustness in low SNR noisy environments. The online experiments show that it can also obtain a good performance in real-time processing conditions.