rConverse: Moment by Moment Conversation Detection Using a Mobile Respiration Sensor

Area of Science:

• Physiological signal processing
• Human-computer interaction
• Speech and language technology

Background:

• Traditional conversation monitoring relies on acoustic sensors, which can be intrusive or limited in noisy environments.
• Breathing patterns are intrinsically linked to physical and cognitive activities, including speech production.
• Analyzing subtle physiological signals offers a non-intrusive alternative for understanding human interaction.

Purpose of the Study:

• To develop and validate a novel method for detecting moment-by-moment conversation episodes using mobile respiration sensor data.
• To establish a robust pipeline for cleaning, screening, and analyzing noisy respiration data at the breath cycle level.
• To compare the performance of the proposed respiration-based system against high-quality audio recording for conversation detection.

Main Methods:

• Development of a comprehensive data cleaning and analysis pipeline for noisy respiration signals.
• Implementation of a Conditional Random Field, Context-Free Grammar (CRF-CFG) based model (rConverse) for classifying respiration cycles.
• Validation using lab-based studies with speech dynamics analysis and field studies with audio ground-truth comparison.

Main Results:

• Individual respiration cycle identification achieved 96.34% accuracy, even during walking.
• The rConverse model demonstrated 82.7% accuracy in speech/non-speech classification and 95.9% accuracy in identifying conversation episodes on lab data.
• In field studies, rConverse achieved 71.7% accuracy in detecting conversation episodes, comparable to high-quality audio recorders (71.9%).

Conclusions:

• Analyzing breathing patterns via mobile respiration sensors is a viable and accurate method for detecting conversation episodes.
• The rConverse system offers a non-intrusive, accurate alternative to acoustic monitoring for conversation analysis.
• This approach has significant potential for applications requiring unobtrusive monitoring of social interactions.