Evaluating automatic creaky voice detection methods

Area of Science:

• Phonetics and Speech Science
• Acoustic Phonetics
• Computational Linguistics

Background:

• Growing research interest in non-modal voice quality, specifically creaky voice.
• Manual annotation of creaky voice is time-consuming, necessitating automatic detection methods.
• Existing automatic methods utilize diverse acoustic cues for creak detection.

Purpose of the Study:

• To compare the performance of three automatic creaky voice detection tools: AntiMode, Creak Detector, and Roughness.
• To investigate the efficacy of combining these tools for enhanced creak detection accuracy.
• To identify strategies for improving automatic creak detection, particularly for large-scale studies.

Main Methods:

• Comparison of AntiMode, Creak Detector, and Roughness algorithms against manual annotation.
• Utilized speech data from 80 Australian English speakers.
• Explored the impact of combining tools and restricting analysis to sonorant segments.

Main Results:

• Combining automatic tools can yield more accurate creak detection than individual tools.
• Restricting analysis to sonorant segments significantly improves automatic creak detection performance.
• Tools performed more consistently on female speech compared to male speech.

Conclusions:

• The study provides practical options for researchers, including combining automatic tools for improved creak detection.
• Optimizing detection via a creak probability threshold sweep is recommended before applying the Creak Detector algorithm.
• Findings support efficient, large-scale research on creaky voice by offering promising automated solutions.