Automatic detection of laughter in conversational speech
- Status
- Finished
- Type
- Master Thesis
- Announcement date
- 06 Dec 2018
- Student
- Witold Łuszcz
- Mentors
- Research Areas
Laughter present in conversational speech is the most common nonverbal occurrence next to respiration. The speaker can express internal emotional states in addition to the lexical-semantic meaning of the utterance. While acoustic features of laughter such as pitch, voicing, and intensity are easy to analyse with tools such as scientific speech analysis software or machine learning algorithms, analysing emotions, which are individual human characteristics, is more challenging.
The aim of this master thesis is to analyze and compare the function, emotions, and acoustic properties of laughter in subjective tests and, at the same time, explore how different functions and emotions of laughter are realized acoustically. Furthermore, the aim is to see whether laughter is perceived differently in one’s own mother tongue than when heard in an unknown language. For that purpose, laughter tokens with a laugh were extracted from two conversational speech corpora of Common Czech and Austrian German, and acoustic analysis and listening experiments were performed with Czech and Austrian participants.
The thesis presents findings from the manual acoustic analysis, auditory-phonetic analysis and from a survey in which the emotions were classified by 33 respondents. Important acoustic features that allow for distinguishing mirthful and embarrassing laughter from polite and derisive laughter have been identified. The relationship between the two emotion classification systems is also shown