Investigation of Segmentation in i-Vector Based Speaker Diarization of Telephone Speech

Abstract

Abstract. The goal of this paper is to evaluate the contribution of speaker change detection (SCD) to the performance of a speaker diarization system in the telephone domain. We compare the overall performance of an i-vector based system using both SCD-based segmentation and a naive constant length segmentation with overlapping segments. The diarization system performs K-means clustering of i-vectors which represent the individual segments, followed by a resegmentation step. Experiments were done on the English part of the CallHome corpus. The final results indicate that the use of speaker change detection is beneficial, but the differences between the two segmentation approaches are diminished by the use of resegmentation.

Experiments with Segmentation in an Online Speaker Diarization System

“…It is assumed that any inaccuracies can be resolved during the later stages of the diarization process, typically by performing a final resegmentation step. This was also confirmed in our recent paper [16], which compared the fixed length approach with SCD-based segmentation using GLR distance. There, we showed that even though the SCD approach led to better initial clusters, the final results of both options after resegmentation were comparable.…”

“…The specific implementation largely follows the descriptions presented in our previous papers [8,16] and a diagram of the main steps can be seen in Fig. 1.…”

“…For these experiments, we chose to use the segmentation algorithms which were previously described in our two recent papers [8,16] in the context of offline speaker diarization. All of these segmentation approaches assume the possibility of their use in online diarization, i.e.…”

“…As the first speaker change detection approach, we used the Generalized Likelihood Ratio (GLR)-based algorithm described in our previous paper [16]. This is a two-pass algorithm, which means that it is not suitable for true online diarization in its current form.…”

“…A smaller number of the most likely speaker change points are found as the local maxima whose topographical prominence exceeds a set threshold. During the second pass, segments above a specific length are split according to the algorithm suggested in [16].…”