Speaker change detection or speaker-based segemtnation is useful and important in many applications, such as transcribing broadcast news or telephone conversations. It usually serves as a preliminary step prior to speech/speaker recognition. Among various methods proposed in the literature. Bayesian Information Criterion (BIC) based method has been widely used. In this paper, we propose to use a different criterion, Minimum Message Length criterion (MML), which is also well known in the statistical community, on speaker change detection problems. MML is an information theoretic criterion that aims to minimize the message length for the description of both model parameters and the data. Previous studies [6] by Oliver etc, in the area other than speech, showed that MML might be a better eriterion than BIC on segmentation problems. We extended their work and applied MML criterion to speaker change detection problems. Experiemtns were carried out on two different types of speech data, and so far, comparable reuslts between BIC and MML have been obtained.
展开▼
