Voice Activity Detection Based on LPCC and Spectrum Entropy
Tang Kun · Telecommunication Engineering · 2010
In order to improve the accuracy of Voice Activity Detection(VAD) in low SNR noisy environments, an algorithm based on Linear Predictive Cepstral Coefficient (LPCC) and energy entropy is proposed. First, the LPCC extracted from the input speech is imported into speech model and noise model, both of which are Gaussian Mixture Model (GMM) separately, to calculate the likelihood ratio of speech to noise. The first-stage VAD decision is made based on the likelihood ratio. Then the spectrum entropy is applied to the second decision-making stage. Finally, a mechanism called Hangover is used to better protect the speech. Experiment results show that the new algorithm can compensate the drawbacks of spectrum entropy method in babble noisy environment. Furthermore, it outperforms the G.729 Annex B under various noisy environments.