부산대학교 도서관

This paper presents a novel method for identifying regions of speech based on the Continues Hidden Markov Model. Probably the most difficult problem faced when using HMMs is that of specifying the model parameters. The Baum-Welch algorithm is a method for estimating model parameters. Unfortunately, for large HMMs, multiplying many probabilities always yields very small numbers that will give underflow errors on any computer. We solve this problem by applying the Baum-Welch algorithm on the small segments of the observed noisy speech vector, based on the statistical approaches. The proposed algorithm estimates model parameters, recursively.Emissions probability distributions are assumed to be Variance Gamma (VG) and Gaussian distributions for speech and non-speech states, respectively.The simulation results show that the proposed Voice Activity Detection (VAD) is able to operate down to -5 dB and in nonstationary environments.