Evalvacija algoritmov za diarizacijo govorcev v zvočnih posnetkih

MILESKI, VANJA (author), Marolt, Matija (mentor)

Segmentacija govorcev (diarizacija) je postopek, ki razdeli zvočni posnetek na odseke glede na identiteto govorcev. Segmentacija govorcev v zvočnih posnetkih nam da odgovor na vprašanje kdo je kdaj govoril. V tem diplomskem delu se posvetimo avtomatični segmentaciji govorcev v različnih zvočnih posnetkih. Pripravimo si testno množico zvočnih posnetkov v slovenščini, ki jih pridobimo iz terenskih posnetkov. Posnetki vsebujejo dva ali več govorcev in zelo pogosto tudi druge zvoke, tišino, prekrivanje med govorcema in podobno. Zanje ročno zgradimo transkripcije za število govorcev in časovni interval govora posameznega govorca, ki bo predstavljal našo resnico (angl. ground-truth). Nad testno množico poženemo vse algoritme za diarizacijo, ki jih ocenjujemo. Napišemo program, ki za vhod vzame rezultate vseh algoritmov, ki imajo različne formate in vrste zastopanja rezultatov, ter jih pretvorimo v enotno obliko. Ocenjujemo natančnost algoritmov in analiziramo, kako dobro delajo v različnih situacijah.

• Language: Slovenian
• Keywords: segmentacija govorcev, diarizacija, terenski posnetki, govor, evalvacija algoritmov
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2015
• PID: 20.500.12556/RUL-72192
• Publication date in RUL: 08.09.2015

Secondary language

• Language: English
• Title: Evaluation of algorithms for speaker diarization in sound recordings

Abstract

Speaker segmentation (diarisation, diarization) is a process that separates the audio clip in sections regarding the identity of the speakers. Speaker diarization in sound recordings answers the question of who spoke when? This thesis is dedicated to the automatic segmentation of speakers in a variety of sound recordings. We prepare a test data of audio recordings in Slovenian, which are obtained from field recordings. The recordings contain two or more speakers and very often they contain other sounds, silence, overlap between the speakers and the like. We manually build transcriptions for the number of speakers in them and the time interval of the speeches for each speaker which will represent our ground-truth. We run all the algorithms for diarization that we evaluate on this test data. We write a program which takes the results from the algorithms as an input which have different formats and different types of representation, and the results are converted to a common format. We evaluate the accuracy of the algorithms and analyse how well they work in different situations.

• Keywords: speaker segmentation, diarization, diarisation, field recordings, speech, algorithm evaluation