Voice Activity Detection (VAD) is considered as an important pre-processing step in speech processing systems such as speech enhancement, speech recognition, gender and age identification. VAD helps in reducing the time required to process speech data and to improve final system accuracy by focusing the work on the voiced part of the speech. An automatic technique for VAD using Fuzzy-Neuro technique (FN-AVAD) is presented in this paper. The aim of this work is to alleviate the problem of choosing the best threshold value in traditional VAD methods and achieves automaticity by combining fuzzy clustering and machine learning techniques. Four features are extracted from each speech segment, which are short term energy, zero-crossing rate, autocorrelation, and log energy. A modified version of fuzzy C-Means is then used to cluster speech segments into three clusters; two clusters for voice and one for unvoiced. After that, three feed forward neural networks are trained to adjust their weights, in which each network represents one cluster. To make the final decision regarding the class type of a given speech segment, the membership degrees of this segment in all clusters along with neural networks' decisions are given to a defuzzification step which finally gives the class type of that segment. The proposed FN-AVAD is tested on the public multimodal emotion database, Surrey AudioVisual Expressed Emotion (SAVEE), and the error rate was 2.08%. The achieved results are comparable to the results achieved by the current published works in the literature.
