How well AssemblyAI Universal-3.5 Pro streaming speech recognition hears dysarthric speech, and what SynapVocal adds on top. SynapVocal is designed for dysarthria and for the speech of people with Parkinson's disease, ALS, cerebral palsy, Down syndrome or stroke; so far it is measured on dysarthria from cerebral palsy or ALS. Measured on the TORGO database (8 speakers with dysarthria, 7 control speakers): a first round of 695 head-microphone utterances, seed 20260917, and a larger pilot with every unique dysarthric sentence. Audio was streamed in real time exactly as the app streams a microphone.
What was tested
Set
Dysarthria sentences
Dysarthria words
Control sentences
Control words
Total
Whole TORGO database (13.7 hours, both microphones)
1321
4253
2841
8137
16552
Speech recognition test (round A and B)
296
120
174
105
695
Suggestion test (Gemini)
296
0
0
0
296
Stability check (streamed three times)
28
4
20
8
60
Larger pilot, speech recognition
682
1318
0
0
2000
Larger pilot, suggestions
681
0
0
0
681
A sentence has more than one word; a word item is a single spoken word. The test sets use one recording per utterance (head microphone). Suggestions were tested on dysarthric sentences only, because that is where speech recognition needs help.
AssemblyAI, dysarthric sentences
0.32WER on 681 sentences
When the Speaker chooses
0.20first suggestion 0.23
Sentences exactly right
418 of 681484 close, 324 without help
Suggestion wait, typical
4.4 sp90 23.1 s
What helps the suggestions
AssemblyAI only
Gemini's first suggestion
Speaker picks best option
AssemblyAI only, 0.32
From the transcript alone, Gemini's first suggestion is slightly worse than AssemblyAI. Hearing the audio makes it better, and earlier sentences from the same person help again. The app uses text + audio + history: the transcript, the sentence audio and earlier sentences from the same person. Letting the Speaker choose lowers the error further.
Where speech recognition struggles
Speakers with dysarthria
Control speakers
Single words are harder than sentences for everyone: one misheard word is a 100% error for that utterance, and there is no context to lean on.
Every speaker is different
Speakers with dysarthria
Control speakers
Dysarthria ranges from mild to severe. For three speakers AssemblyAI is already nearly perfect; for four it misses more than half the words. An average hides this.
First round: text only
Text used without the Speaker choosing
Speaker chooses
AssemblyAI only, 0.40
Neither a context prompt for AssemblyAI nor a Gemini suggestion on its own is more accurate than AssemblyAI alone. Letting the Speaker pick among the options is what lowers the error. Differences under 0.03 are within run-to-run variation.
Suggestion speed
Gemini 3.7 Flash (used)
Gemini 3.8 Flash
Quality was the same for both models (best-choice WER 0.35 and 0.35). 3.7 Flash is used because its slow calls are much less slow.
Are the numbers stable?
Speakers with dysarthria
Control speakers
Control speakers came back identical every time. Dysarthric results moved by up to 0.033, from a few hard utterances the model guesses differently each run.
Limits. TORGO sentences are read aloud and many are well known, so free conversation will be harder. The "Speaker chooses" figures assume the Speaker recognises their own sentence; they are an upper bound, not a user study. Word error rate is corpus-level after removing case and punctuation.
Speech recognition: AssemblyAI Universal-3.5 Pro streaming. Suggestions: Gemini 3.7 Flash. Data: Rudzicz, F., Namasivayam, A.K., Wolff, T. (2012). The TORGO database of acoustic and articulatory speech from speakers with dysarthria. Language Resources and Evaluation, 46(4), 523 to 541. Used for evaluation only; no audio is hosted here. Scripts and raw results.
