Improvement of SVM-based voice activity detection via sparse coding

Voice activity detection (VAD) can be considered as a binary classification problem and solved using the support vector machine (SVM). This paper presents a robust approach to improve the performance of conventional SVM based VAD methods. To this end, we first generate sparse representations by using a speech dictionary learned from clean speech, and derive some kind of audio features from the sparse representations. Then, we design a SVM to detect speech region and non-speech region based on these features. Experiments show that the proposed approach for noise-robust feature extraction further improves the performance of SVM based VAD methods especially in low SNR noisy environments.