SEP-28k: A Dataset for Stuttering Event Detection From Podcasts With\n People Who Stutter

Le résumé fourni par la source

The ability to automatically detect stuttering events in speech could help\nspeech pathologists track an individual's fluency over time or help improve\nspeech recognition systems for people with atypical speech patterns. Despite\nincreasing interest in this area, existing public datasets are too small to\nbuild generalizable dysfluency detection systems and lack sufficient\nannotations. In this work, we introduce Stuttering Events in Podcasts\n(SEP-28k), a dataset containing over 28k clips labeled with five event types\nincluding blocks, prolongations, sound repetitions, word repetitions, and\ninterjections. Audio comes from public podcasts largely consisting of people\nwho stutter interviewing other people who stutter. We benchmark a set of\nacoustic models on SEP-28k and the public FluencyBank dataset and highlight how\nsimply increasing the amount of training data improves relative detection\nperformance by 28\\% and 24\\% F1 on each. Annotations from over 32k clips across\nboth datasets will be publicly released.\n

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.