AUTHOR=Barrett Liam , Tang Kevin , Howell Peter TITLE=Comparison of performance of automatic recognizers for stutters in speech trained with event or interval markers JOURNAL=Frontiers in Psychology VOLUME=Volume 15 - 2024 YEAR=2024 URL=https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1155285 DOI=10.3389/fpsyg.2024.1155285 ISSN=1664-1078 ABSTRACT=Automatic recognition of stutters (ARS) and their type from audio recordings should improve objective analysis of such speech and promote use of assistive technologies by people who stutter. To date, ARS procedures have had mixed success which may be due, in part, to the segmentation types used to prepare training and test materials. The segmentation types are either event-based (segments that are stuttered or fluent are delimited), or interval-based (fixed-length segments are obtained from continuous speech). Performance may also be improved if long stretches of speech that extend beyond the segments are used as they allow linguistic information to be used. Interval segments may include such linguistic information, but this is absent from event segments. Whilst this suggests intervals would be preferred over events, any fixed-length interval with a stuttering event that lasts for less time than the interval's duration contains both stuttered and fluent speech. This would make ARS more challenging than when event-segments are used that contain only stuttered or only fluent speech. Given the pros and cons for the two segmentation approaches, this study addressed whether interval-, or event-procedures are preferred for ARS. Machine learning models were trained and evaluated on a recent interval-based stuttered speech corpus (Bayerl et al., 2022a) and an established event-based corpus (Howell et al., 2009) which was also formatted as intervals for this study. Overall, event-based training led to better model performance (area under the curve). The implications for development of ARS systems for application to typical and disordered speech are discussed.