Identifying languages in a novel dataset: ASMR-whispered speech

Area of Science:

• Neurology
• Linguistics
• Computer Science

Background:

• Autonomous Sensory Meridian Response (ASMR) is a popular phenomenon characterized by tingling sensations triggered by specific auditory or visual stimuli.
• Despite its popularity, ASMR remains largely unexplored due to the lack of accessible, open-source datasets for research.
• This gap hinders the development of computational models for analyzing ASMR-related content.

Purpose of the Study:

• To introduce the ASMR Whispered-Speech (ASMR-WS) database, a novel resource specifically designed for ASMR research.
• To facilitate the development of unvoiced Language Identification (unvoiced-LID) systems capable of recognizing languages within ASMR stimuli.
• To provide a foundation for further scientific investigation into the ASMR phenomenon.

Main Methods:

• The ASMR-WS database was created, comprising 38 videos totaling over 10 hours of whispered speech.
• The database includes recordings in seven languages: Chinese, English, French, Italian, Japanese, Korean, and Spanish.
• Baseline unvoiced-LID models were developed using Convolutional Neural Networks (CNNs) and Mel-frequency cepstral coefficients (MFCCs) on 2-second segments.

Main Results:

• The best-performing model achieved an unweighted average recall of 85.74% and an accuracy of 90.83% on the seven-class unvoiced-LID task.
• These results demonstrate the feasibility of identifying languages from whispered speech within the ASMR context.
• The study provides initial performance benchmarks for future research.

Conclusions:

• The ASMR-WS database is now publicly available to the research community, promoting further study of ASMR.
• Future research will explore the impact of speech sample duration on unvoiced-LID system performance.
• The database and baseline findings offer a significant contribution to the under-explored field of ASMR research.