Voice activity detection based on noise classification and dictionary selection

The performance of current voice activity detection (VAD) methods drops substantially in noise condition. To solve this problem, a new VAD method based on sparse coding was proposed. In the training, this method learns a dictionary for speech signals and each possible noise; in the testing, this method first identifies environmental noise types, and then concatenates the speech dictionary and corresponding environmental noise dictionary to be a large dictionary for sparse decomposition, and finally uses the representation over speech dictionary to make speech and non-speech classification. Since making using of noise classification, this method can select noise dictionaries. In addition, this method makes use of out-set recognition of noises, which can recognize new noisy and train models for them. Experiments results show that the proposed method is more robust than traditional methods.