SphereDiar: Tehokas puheen diarisointijärjestelmä kokousäänitteitä varten
Year: 2019
Type: dissertation
Abstract: The objective of speaker diarization is to determine who spoke and when in a given audio stream. This information is useful in multiple different speech related tasks such as speech recognition, automatic creation of rich transcriptions and text-to-speech synthesis. Moreover, speaker diarization can also play a central role in the creation and organization of speech-related datasets. Speaker diarization is made difficult by the immense variability in speakers and recording conditions, and the unpredictable and overlapping speaker turns of spontaneous discussion. Especially diarization of meeting data has been very challenging. Even the most advanced speaker diarization systems still struggle with this type of data. In this thesis, a novel speaker diarization system, named SphereDiar and designed for the diarization of meeting data, is proposed. This system combines three novel subsystems: the SphereSpeaker neural network for speaker modeling, a segmentation method named Homogeneity Based Segmentation and a clustering algorithm Top Two Silhouettes. The system harnesses up-to-date deep learning approaches for speaker diarization and addresses the problem of overlapping speech in this task. Experiments are performed on a dataset consisting of over 200 meetings. The experiments have two main outcomes. Firstly, the use of Homogeneity Based Segmentation is not vital for the system. Thus, the configuration of SphereDiar can be simplified by omitting segmentation. Furthermore, SphereDiar is shown to surpass the performance of two different state-of-the-art speaker diarization systems.
Source: Aaltodoc (Aalto University)
Author: Kaseva, Tuomas
Institution: Aalto University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7124120435
