Whereas the final decade has witnessed important developments in Computerized Speech Recognition (ASR) methods, efficiency of those methods for people with speech disabilities stays insufficient, partly as a result of restricted public coaching information. To bridge this hole, the 2025 Interspeech Speech Accessibility Venture (SAP) Problem was launched, using over 400 hours of SAP information collected and transcribed from greater than 500 people with numerous speech disabilities. Hosted on EvalAI and leveraging the distant analysis pipeline, the SAP Problem evaluates submissions based mostly on Phrase Error Fee and Semantic Rating. Consequently, 12 out of twenty-two legitimate groups outperformed the whisper-large-v2 baseline by way of WER, whereas 17 groups surpassed the baseline on SemScore. Notably, the highest group achieved the bottom WER of 8.11%, and the very best SemScore of 88.44% on the similar time, setting new benchmarks for future ASR methods in recognizing impaired speech.
- † College of Illinois Urbana-Champaign
- ‡ Microsoft
- § Amazon
- ‡‡ Meta
- §§ CDLI
- ¶¶ Inha College
- ††† KAIST
