Robust Voice Activity Detection Based on Feature Fusion and Recurrent Neural Network

Voice activity detection (VAD) is an essential pre-processing step for some of speech tasks as automatic speech recognition (ASR), speech enhancement, speech recognition and speaker recognition. It is applied to classify an input signal into two main parts: non-speech and speech parts. In this paper, VAD model is proposed to classify speech and non-speech frames. it splits input signal into frames by hamming window. Three feature extraction methods are utilized to prepare data for classification phase: Mel-frequency Cepstral Coefficients (MFCC), log filter banks and spectral subband centroid. The retrieved features are fused and taken as input to the classification phase. It consists of three Recurrent Neural Network (RNN) with 256 neurons and fully connected layer with two neurons. The proposed model has been tested and trained using a dataset that consisting of 719 audio files. The experimental and validation results proved the efficiency of feature fusion and RNN in recognizing speech and non-speech frames in the input signal, where it achieved training and validation accuracy reached to 95% and 88% respectively.