Speech/music discrimination for radio broadcasts using a hybrid HMM-Bayesian Network architecture

This paper presents a speech/music discrimination scheme for radio recordings using a hybrid architecture based on a combination of a Variable Duration Hidden Markov Model (VDHMM) and a Bayesian Network (BN). The proposed scheme models speech and music as states in a VDHMM. A modified Viterbi algorithm for the computation of the observations´ probabilities at each state is proposed. This is achieved by embedding a BN, that outputs to the HMM the required probability values. The proposed system has been tested on audio recordings from a variety of radio stations and has exhibited an overall performance close to 95%.