8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


A DP Algorithm for Speaker Change Detection

Michele Vescovi (1), Mauro Cettolo (2), Romeo Rizzi (1)

(1) Universita degli Studi di Trento, Italy
(2) ITCirst, Italy

The Bayesian Information Criterion (BIC) is a widely adopted method for audio segmentation; typically, it is applied within a sliding variable-size analysis window where single changes in the nature of the audio are locally searched. In this work, a dynamic programming algorithm which uses the BIC method for globally segmenting the input audio stream is described, analyzed, and experimentally evaluated. On the 2000 NIST Speaker Recognition Evaluation test set, the DP algorithm outperforms the local one by 2.4% (relative) F-score in the detection of changes, at the cost of being 38 times slower.

Full Paper

Bibliographic reference.  Vescovi, Michele / Cettolo, Mauro / Rizzi, Romeo (2003): "A DP algorithm for speaker change detection", In EUROSPEECH-2003, 2997-3000.