Speech Recognition for People with Atypical Speech

Speech recognition systems are commonly trained on large collections of speech, but the speech of people with dysarthria and other atypical patterns may be underrepresented. Differences in articulation, timing, voice quality, and prosody can increase recognition errors. These are not simply “bad speech”; they reflect variation that mainstream systems may not model well. Personalization is one research approach. Project Euphonia research has explored models for non-standard speech, and Google’s Project Relate Android beta offers Listen, Repeat, and Assistant functions for some users. Google’s current Project Relate page says it is not accepting new users while existing users can continue to access models. Voiceitt describes a separate product that asks users to record phrases and can support speech-to-text or synthesized output in specified workflows. These product details and availability can change; check current vendor documentation. Research results must be read in context. A 2019 paper on personalized ASR reported relative word-error-rate improvements for study groups speaking dysarthric or accented English, using limited data and message-bank test phrases. That does not establish performance on spontaneous conversation, all conditions, languages, or microphones. A 2025 conversational-speech study with 27 participants also underscores the need to evaluate real conversational language, not only prompted phrases. Users should be able to correct transcripts, train or opt out as they choose, and keep an alternate communication method. Evaluate performance on the person’s own words, names, noisy environments, and intended tasks. Discuss data retention and sharing before recording voice samples. Recognition can support access but should not be presented as guaranteed communication or a replacement for AAC, speech-language support, or human listeners.