A new study led by Prof. LI Hai's team from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences has evaluated the reliability of speech acoustic features across common consumer-grade mobile devices.

Published in Behavior Research Methods, the study concludes that frequency-related speech features, such as fundamental frequency and cepstral peak prominence, exhibit high reliability across various devices.

Remote speech-based cognitive assessments are an emerging tool that leverages internet-connected devices to analyze speech and evaluate cognitive function. This technology offers a convenient, accurate, and non-invasive way to monitor cognitive decline in older adults and mental health in adolescents. However, ensuring consistent results across different devices and repeated measurements has been a challenge, and little research has been done to address this issue in remote settings.

To address this challenge, the research team assessed the cross-device and test-retest reliability of speech acoustic features using four common consumer-grade devices: digital voice recorders, laptops, tablets, and smartphones. Their research found that frequency-related features, such as fundamental frequency and cepstral peak prominence, demonstrated high reliability across different devices and measurements, making them suitable for remote assessment. In contrast, more complex features, like syllable rate and regularity, showed lower reliability.

The study also emphasizes the need for standardized protocols for data collection and analysis to improve the reliability of remote speech assessments. Additionally, the researchers suggested that improving algorithms could help consumer-grade devices handle complex acoustic features more accurately in noisy environments. 

These findings provide crucial data to advance remote speech-based assessment technologies, which hold significant promise for applications in cognitive health monitoring, especially for aging populations and mental health monitoring in adolescents.

Cross-device reliability of Acoustic Features in Remote Speech Assessment (Image by YANG Lizhuang)