Accessibility
Rubil helps a specific group of people a great deal, and there is another group it has not earned the right to claim yet. This page says which is which.
Where Rubil helps today
If typing is painful or difficult for you and speaking is not, Rubil works for you now, with no caveats and no special configuration. That includes people managing repetitive strain injury, chronic pain, arthritis, hand or arm motor impairment, and anyone computing one-handed.
The reason is simple. For this group the speech itself is typical, so speech recognition performs at its normal accuracy. Typing is the barrier, and removing typing is the whole product. Every accuracy claim we make elsewhere on this site holds here.
One thing that matters in academic and professional settings: Rubil never generates content on your behalf. It formats what you said and stops. Where an assessor or an employer scrutinises AI tools for authorship, that distinction is the point. The ghostwriter philosophy explains it in full.
Where Rubil has not been validated
If your speech is affected by dysarthria, from Parkinson's, cerebral palsy, ALS or stroke, or if you stutter, we have not tested Rubil with you and we are not going to claim it works. If you would like to help us find out, we would genuinely welcome that, and the door is open to anyone willing: people who want to test it themselves, speech-language pathologists, accessibility organisations, researchers, or carers and family who see this up close. Write to [email protected]. We will tell you exactly what testing would involve before you agree to anything, you set the terms, and you can stop at any point. Nothing gets published about your speech without your say-so.
Rubil uses Whisper-class speech recognition. Published research on that class of recognition, not on Rubil specifically, reports steep degradation on these speech patterns:
- A study of 211 speakers with Parkinson's-related dysarthria found a baseline recogniser that transcribes typical speech at 3.4% word error rate transcribed dysarthric speech at 36.3%, roughly ten times worse. Fine-tuning on dysarthric data improved it to 23.7%, still about seven times the typical rate.
- Whisper evaluations on standard dysarthric corpora report average word error rates of 12.7% to 21.5% on sentence material, and worse on isolated words.
- On disfluent speech, one controlled study reports word error rate rising from 3.1% to 19.8%. Sound repetitions and blocks, the disfluency types most characteristic of stuttering, degrade it most.
Every figure above is linked at the bottom of this page. We would rather hand you the research than a reassurance.
Why we will not simply say “give it a try”
There is a specific failure mode here that makes a soft answer dishonest. Benchmark work on recognition pipelines for dysarthric speech finds that Whisper often hallucinates words, producing fluent but incorrect transcriptions.
A fluent-but-wrong transcript flowing into Rubil's formatting stage produces confident, well-formatted text that you never said. That is the exact harm the entire product exists to prevent. For this group, Rubil could look like it is working while breaking its central promise, and that is worse than not working at all.
Rubil's Glossary cannot rescue this either. It operates at the formatting stage and cannot recover a word that was misheard. The formatting layer is also forbidden from guessing at what you meant, because guessing is generation.
What would change this
What fixes recognition for atypical speech is acoustic personalization. Google's Project Euphonia showed personalized models improving word error rate by up to 85% over off-the-shelf models, with as little as three to four minutes of a person's own speech. Rubil does not do this today.
Before we claim anything for this group, we intend to run a small, consent-based trial supervised by a speech-language pathologist, with volunteers who have mild dysarthria or a mild stutter, measuring the word-level difference between what a person intended and what Rubil produced. We will publish the result either way, including the limits.
If you are a speech-language pathologist or an accessibility organisation and this is your area, we would rather hear from you before we make claims than after. [email protected].
Formal conformance status
Rubil has not been through a formal WCAG audit, and there is no VPAT for it yet. We are not going to imply otherwise while that is true. If you need either for a procurement or institutional review, write to us and we will tell you honestly where it stands.
If you hit an accessibility barrier in the extension, the Mac app, or this website, tell us at [email protected]. It is read by the people who build the product.
Sources
Every figure on this page comes from one of these. None of them is a study of Rubil. They describe the class of speech recognition Rubil is built on.
- Speech Accessibility Project, Parkinson’s dysarthria study (University of Illinois / JSLHR)
- Fine-Tuning ASR for People with Parkinson’s (arXiv 2409.19818)
- Bridging ASR and LLMs for Dysarthric Speech, hallucination finding (arXiv 2508.08027)
- AAC evaluation on the TORGO corpus (arXiv 2411.00980)
- Lost in Transcription: ASR bias against disfluent speech (arXiv 2405.06150)
- Whisper disparities on stuttered speech (Interspeech 2025)
- Project Euphonia, personalized ASR models (Google Research)
Literature reviewed July 20, 2026. If you find a study that changes the picture, send it to us.