Short Description

When two speakers are speaking with each other, they are taking turns in speaking. Depending on the context of the conversation and on the relationship between the speakers, conversations may look very differently from the dynamics in how speakers contribute to the conversation. Speakers might speak strictly one after another, or they might speak also frequently at the same time (e.g., if they are good friends); they might ask concrete questions and answer to them (e.g., in the context of information seeking), or they might build up a story collaboratively (e.g., when remembering something that they experienced together in the past). The aim of this project/thesis is to analyze the characteristics of different conversational dynamics and to use use these characteristics (acoustic, semantic and rythmical features) to automatically classify such patterns. The figure shows a 100 sec. long stretch of a conversation with different turn-taking functions (as annotated manually), their sequential occurrence in time for two befriended speakers (028F and 008M) that are talking to each other in a casual way.

Teams of 2-3 students are welcome!

Your Tasks:

• Extract acoustic and semantic features from the Grass Speech Database
• Train different types of classifiers suitable for small data sets (e.g., Random Forests)
• Use Shapely tools to investigate the features’ importance
• Analysis of the classification and reporting your results (thesis writing)

Your Profile

• basic knowlegde of sound engineering and/or speech communication
• good knowledge of programming (e.g., Python)

Contact:

Barbara Schuppler (b.schuppler@tugraz.at)