National Technical University of Athens (NTUA)
Νέες μπεϋσιανές τεχνικές ομαδοποίησης με εφαρμογές στην αυτόματη δεικτοδότηση ομιλητών σε αρχεία ήχου
Abstract
dc:descriptionThis thesis focuses on the problem of segmentation and clustering of audio files to speakers, of problem termed in literature as speaker diarization. It is considered as a central problem, since many applications that are related to speech technologies require it as a preprocessing step. They require an algorithm that is capable of estimating in a computationally efficient way (a) the number of speakers and (b) the time segments that each of the speakers is active. Compared to other clustering and classification tasks, speaker diarization exhibits a pair of special characteristics that makes it both attractive and hard-to-tackle; the lack of knowledge of both the number of speakers and their identity. Three are the main contributions of the thesis: (a) Our first contribution is the development of a probabilistic measure of discrepancy between two speech segments. This discrepancy aims to estimate the posterior probability of the segments to belong to different speakers. The proposed model is capable of combining an unlimited number of binary weak classifiers, each of which should be considered as a combination of a feature space, a statistical model, a statistical divergence and a threshold. (b) The second contribution is a redefinition of one of the most frequently used approaches to speaker diarization, namely the Bayesian Information Criterion (BIC). By examining the Bayesian rationale for BIC, we show that both of its current versions (the global and the local) are suboptimal for speaker diarization. Using the implied priors of BIC, we we propose a new version, the segmental-BIC, that leads to a significant increase in diarization accuracy. (c) Our third and final contribution is the examination of the potentials that the mean shift algorithm offers to the problem of speaker diarization. This algorithm is highly used in the image processing and computer vision and has been established a milestone in nonparametric segmentation. We show that it can be used to tackle more general clustering tasks, where the entities lie on non-Euclidean spaces, like those of statistical parametric models of exponential families. Using elements of Information Geometry and a Bayesian framework, we adapt the original algorithm and show that is capable of increasing the diarization accuracy when compared to the standard hierarchical clustering approach.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Stafylakis, Themos
- Σταφυλάκης, Θέμος
Subjects
dc:subject × 14- Ομαδοποίηση ομιλητών
- Μπεϋσιανή στατιστική
- Γεωμετρία πληροφορίας
- Μπεϋσιανό κριτήριο πληροφορίας
- Αλγόριθμος μετατόπισης του μέσου
- Speaker clustering - diarization
- Bayesian statistics
- Information geometry
- Bayesian information criterion
- Mean shift algorithm
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/25570