Estimating Speech Duration by Measuring the Abdominal Movement Using a Barometric Sensor
Rintaro Katagiri, Yutaka Arakawa, Yugo Nakamura
TL;DR
This study addresses privacy concerns in measuring daily speech by proposing a barometric-sensor-based abdominal motion device that estimates speech duration without microphones. The authors collect data from 10 subjects under speech and no-speech conditions, extract statistical and FFT features from abdominal motion, and train multiple classifiers, finding random forest to perform best with up to 94.6% accuracy on lab data. However, performance degrades in real meetings due to posture changes, leading to overestimation of speech duration. The work demonstrates a privacy-preserving approach to quantify speech production and highlights posture as a key factor to overcome in future iterations for real-world applicability.
Abstract
Measuring the amount of speech production in daily life is important for understanding communication in organizations and identifying mental disorders. However, measuring the amount of speech production can be problematic in terms of privacy. We observed the whole body condition during speech and noted that the abdomen strains during speech production.Therefore, we developed a less uncomfortable, inflatable abdominal motion measurement device using a barometric sensor to measure speech production indirectly. We measured speech production in 10 subjects and created a speech discrimination model using machine learning. However, the estimated speech duration in an actual meeting using this model was much longer than the actual duration. We found that the wearer's posture significantly affects the accuracy of the speech discrimination model developed in this study. We plan to improve the abdominal motion measurement device to minimize the effect of posture and achieve more accurate speech production measurement.