انجمن کامپیوتر ایران

One of the most important phases of speaker indexing is speaker clustering which aims to find the number of speakers in a speech document and merge the speech segments corresponding to a single speaker. The most critical source of problem in speaker clustering is the speech segments duration which may be so short that proper segment modeling becomes hard to achieve. An alternative suggestion in these situations is to adapt global models with new data instead of building the speaker models from the ground. In this paper we investigate two adaptation techniques in eigen-voice space for improving clustering performance especially for shorter speech utterances. These techniques were embedded in a clustering framework and evaluated on a set of domestic conversational speech. We have also compared the proposed methods with some other known techniques. The experiments show a considerable improvement in speaker clustering performance.

In this paper we study combinatorial segment visibility, and show how to compute and maintain it as the observer moves in the plane. There are n non-overlapping convex objects in the plane, and we have a segment observer among them. We first consider static case of the problem, in which the observer and objects are static, and then we study dynamic case of the problem, in which the observer can move among obstacles.