Synthesizing naturalistic laughter: An exploratory study on modeling voiced laughter with speech synthesis techniques

• Main Authors: Cagampan, Bernadyn R., Ng, Henry O., Panuelos, Kevin Matthew C.H., Uy, Krystyn Kaizzle S.
• Format: text
• Language: English
• Published: Animo Repository 2013
• Subjects: Speech processing systems; Speech synthesis; Laughter; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/10829
• Institution: De La Salle University

This study focuses on the synthesis of naturalistic voiced laughter, and attempts to address the wide gap present in applications that involve synthetic agents. This gap lies in the interactions between human and these agents, which can in part be filled through the emulation and expression of paralinguistic sounds such as laughter. Most agents speak through a synthesized voice, but inserting a prerecord laughter sound in between sentences proved to score low in participation tests (Trouvain & Schroder, 2004), thus making for a compelling reason to pursue computer-generated laughter. This involves the analysis of a set of acoustic features including, but not limited to, pitch and MFCCs present in voiced laughter, and consequently the synthesis of laughter using concatenative diphone synthesis, articulatory synthesis and hidden Markov model-based statistical parametric synthesis techniques. With this in mind, the goal is to generate laughter that is perceived as acceptable and natural by evaluators. This can be validated through subjective evaluation tests where an evaluator determines the synthesized laughter from a set of clips. The results of this work show that while evaluators are primarily able to identify natural laughter from synthesized laughter, there is much doubt and little agreement on whether or not these clips were even truly natural or not. Aside from articulatory synthesis-which was consistently rated lowly- the concatenative diphone synthesis and statistical parametric synthesis techniques proved quite effective in synthesizing laughter that was rated to be even more naturalistic than samples from a spontaneous laughter database. Differences between male and female evaluator groups were found and identified through the use of decision tree models and are used to identify how certain features may have influenced the evaluation score.