On hierarchical clustering for speech phonetic segmentation

In this paper, we face the problem of phonetic segmentation under the hierarchical clustering framework. We extend the framework with an unsupervised segmentation algorithm based on a divisive clustering technique and compare both approaches: agglomerative nesting (Bottom-up) against divisive analysis (Top-down). As both approaches require prior knowledge of the number of segments to be estimated, we present a stopping criterion in order to make these algorithms become standalone. This criterion provides an estimation of the underlying number of segments inside the speech acoustic data. The evaluation of both approaches using the stopping criterion reveals good compromise between boundary estimation (Hit rate) and number of segments estimation (over-under segmentation).