The Relationship Between Surprisal and Prosodic Prominence in Conversation Reflects Intelligibility-Oriented...

Area of Science:

• Computational Linguistics
• Psycholinguistics
• Speech Communication

Background:

• Conversation is a multimodal exchange of information.
• Word duration is linked to predictability, but the cause (speaker vs. listener-oriented) is debated.
• Listener-oriented intelligibility pressures may influence speech production.

Purpose of the Study:

• To investigate the relationship between word predictability (surprisal) and acoustic variables in natural conversation.
• To determine the extent to which intelligibility-based principles impact conversational speech.
• To analyze prosodic features (duration, intensity, pitch) in relation to surprisal.

Main Methods:

• Utilized the GPT-2 large language model to calculate surprisal (unpredictability).
• Analyzed the CANDOR corpus of naturalistic spoken English conversations.
• Examined prosodic features like duration, intensity, and pitch, and listener backchannels.

Main Results:

• GPT-2 surprisal significantly predicted higher word duration and maximum pitch.
• Surprisal also predicted pitch range, with less clear effects on intensity.
• Listener backchannels correlated with spikes in speaker word surprisal.
• Context window size affected model fit differently for pitch versus other variables.

Conclusions:

• Findings support intelligibility-based accounts of language production.
• Speech production appears sensitive to the need for successful communication.
• Conversational dynamics are influenced by listener comprehension needs.