Does speech recognition work for aphasic speech?
First, two terms
Word error rate (WER) is the standard measure of transcription accuracy: the number of words the system got wrong, expressed as a percentage of the words actually spoken. Errors come in three kinds, all counted: substituting a wrong word, deleting a word, or inserting one that was never said. Because insertions are counted, WER can exceed 100 percent when a system adds more words than it heard.
Pronunciation assessment is a different kind of service. Instead of asking "what did this person say?", it takes the audio and the text of what they were supposed to say, and scores how closely the two match. It was built for language learning. The output is a score, not a transcript.
One more, because the numbers below hinge on it: a phonological paraphasia is saying a word with the wrong sounds in it, like "cuv" for "cup". The person has retrieved the right word; the sounds came out wrong. This is common in some kinds of aphasia and it is exactly the case where machines and humans disagree most.
Why general speech-to-text fails on aphasic speech
Every mainstream transcription system is trained on typical speech. The published numbers on atypical speech are consistent and not close.
On the AphasiaBank corpus, Sanguedolce et al. (Interspeech 2024) measured zero-shot Whisper at 40.3 percent WER for the tiny model, 38.1 percent for base, 33.8 percent for small and 35.8 percent for medium. A 2026 evaluation framework published in Aphasiology put Whisper large-v3 at 38.4 percent WER for people with aphasia against 25.0 percent for controls on the same task.
Cloud services fare no better on isolated words. A comparison of ASR platforms on aphasia assessment batteries (PMC9863375) tested Azure and Google speech-to-text on isolated-word naming and repetition from 12 Mandarin-speaking patients with aphasia. Both scored under 31 percent word accuracy. The same systems managed roughly 88 percent (Azure) and under 75 percent (Google) on healthy speakers.
Single-word clips are the worst case of all, and this is the finding families most need to know. Davudova et al. (2025) ran Whisper small and medium on the SONIVA-Naming set: 87 patients, 2,609 single-word picture-naming trials. WER came out between 97.6 and 100.5 percent. The target word was detected in 41 to 45 percent of patient trials, and, tellingly, in only 36 to 40 percent of healthy trials. Whisper-family models are trained on 30-second segments, so given one word they pad it with invented context. This failure is not about aphasia at all; it is about the clip being short.
Related: Koenecke et al. ("Careless Whisper", FAccT 2024) found Whisper hallucinated in 1.7 percent of aphasia clips versus 1.2 percent of control clips, a statistically significant difference (p = 0.019). The driver was silence. In their aphasia audio, 41 percent of the recording was silent, against 15 percent for controls, and pauses are what these models fill with invented text. Long pauses are a defining feature of how many people with aphasia speak.
The pattern generalises past aphasia. On the TORGO dysarthria corpus, a 2025 evaluation of Whisper large-v3, Deepgram Nova-3, AssemblyAI, GPT-4o, GPT-4o-mini and Gemini 2.5 Pro and Flash found the leaders at 1 to 2 percent WER on mild speech and every system above 51 percent WER on severe speech. The large multimodal models held no advantage over conventional ASR.
Fine-tuning and personalisation work, but you cannot buy them
The research shows the problem is tractable. It is just not available to an app.
Sanguedolce et al. fine-tuned Whisper medium on seven hours of the SONIVA corpus and cut WER from the mid-30s to 21.5 percent on AphasiaBank and 14.7 percent on unseen SONIVA data. Davudova et al. fine-tuned on naming clips and got WER to 26.4 percent with target-word detection accuracy of 0.92, up from 0.45. On their 524-trial test, that model produced 2 false positives against 39 false negatives, meaning it erred heavily toward missing a correct attempt rather than crediting a wrong one. None of those weights are public; the SONIVA data is released to academic researchers only.
Google's Project Euphonia personalised models are the most striking result. Trained per speaker on a corpus of 430 speakers with disordered speech, they took severe speakers on short phrases from a median 89 percent WER down to 13 percent, an 85 percent relative reduction, and beat human transcribers for 80 percent of speakers. But Project Relate is an Android app for an individual's own use, with no API a web app can call, and aphasia is not one of the etiologies in that corpus.
Verifying a known word is a much easier problem
Transcription asks an open question with roughly 50,000 possible answers. Verification asks a closed one: here is the audio, here is the word "cup", is that what she said? Three things make it easier.
- The search space collapses. A recogniser restricted to "cup", its accepted alternates, and a catch-all "unknown" cannot return "cub" or "come" as a competing answer.
- You can compare sounds to a reference instead of decoding to text, which gives you a partial score rather than a yes or no. A paraphasia shows up as a partial match, which is the "nearly" band you actually want.
- You can tune the threshold for one person, which a general transcription service will never do for you.
The strongest published result here is NUVA (Barbera et al. 2021, Computer Speech and Language). It computes phone-level probabilities from a model trained on healthy British English, aligns them against reference recordings of the same word, and classifies. It was tested on 8 British people with aphasia, 220 picture-naming items, 1,760 recordings, each labelled by speech and language therapists.
Per-patient accuracy ranged from 83.6 to 93.6 percent, with a ten-fold mean of 89.5 percent. An adapted version produced 6.6 percent false positives and 3.9 percent false negatives. On the same clips, Google Cloud speech-to-text wrapped as a verifier scored 88.2 percent against NUVA's 91.3 percent, and NUVA ran 1.4 times faster. The code is open source.
One detail in that study matters for interpreting every number on this page: the therapists labelled phonological errors as incorrect. A verifier trained on those labels will reject "cuv" for "cup". Whether that is the right behaviour in a home practice tool is a design choice, not a technical fact.
Two other closed-set results point the same way. Apple's Latent Phrase Matching (Interspeech 2023), tested on 32 speakers with dysarthria across 50 phrases with 5 enrolment recordings each, scored 0.84, 0.78 and 0.70 accuracy for mild, moderate and severe speech, where commercial ASR on the same data scored 0.78, 0.44 and 0.02. On single words from the EasyCall corpus, ASR scored 0.55 and the matching approach 0.65. And a 2026 single-case report trained a small network on 1,120 recordings of 13 words from one person with global aphasia and dysarthria: 72.7 percent accuracy on a 13-way choice, against 56.8 percent for familiar rehabilitation professionals attempting the same task by ear.
What pronunciation assessment does, and what it does not
The closest commercially available thing to a verifier is Azure Speech pronunciation assessment in scripted mode. You send the audio and a reference text, which can be a single word, and get back an accuracy score from 0 to 100 at the whole-utterance, word, syllable and phoneme level, plus an error type per word: none, mispronunciation (triggered below a word accuracy of 60), omission or insertion.
The phoneme-level breakdown is the genuinely useful part. For "cuv" said for "cup", you would expect the k and the vowel to score near normal and the final p to score low. That is more information than any transcript gives you.
The limits are stated by Microsoft itself, and they are serious ones here. The documentation reports over 0.5 Pearson correlation with human judges, describes the scores as "relative indicators, not absolute measures", notes that repeated utterances of the same word vary, and flags that the model scores strictly on slow, careful, hyper-articulated speech. Slow and effortful is normal speech for many people with aphasia, so scores will run systematically low. The transparency note lists language learning as the intended use and asks for a human in the loop for anything formal.
Most importantly: there is no published study testing Azure pronunciation assessment on aphasic speech. Anyone who tells you an accuracy figure for it on aphasia is extrapolating. The nearest real data point is the under-31-percent word accuracy that Azure's standard transcription scored on aphasic single words, which is precisely why its transcript should not be used alone.
Browser speech recognition is weaker still. The Web Speech API's grammar list is deprecated and has no effect on recognition, and the newer contextual biasing feature requires on-device recognition, which Chrome on Android does not ship. On an Android tablet you therefore get an unbiasable cloud recogniser with no confidence value you can trust.
What has happened when apps have actually tried this
iTalkBetter (UCL) is the proof that it can work. In a phase II item-randomised trial, 27 people with chronic post-stroke aphasia practised independently while NUVA classified every attempt in real time and drove the cue hierarchy. They averaged 45 hours of practice over 6 weeks, 17,538 trials each. Trained items improved by 13 percentage points, untrained items did not change, gains held at 3 months, and the researchers observed grey-matter increases on MRI. The paper reports no user complaints about misclassification.
VoiceAdapt (2025) is the cautionary half. In a virtual randomised trial with 37 completers doing 19.5 hours over 5 weeks on a tablet with ASR scoring, trained-item naming improved by 13 points (d = 0.9) and the Boston Naming Test by 1.6 points (d = 0.3). One participant withdrew because of frustration with the speech recognition, and the authors explicitly flag that ASR is "not always accurate, particularly given the common co-occurrence of accompanying motor speech impairments".
A 2026 review cites a small iPad study in which an app's speech recognition agreed with human raters about 80 percent of the time and users improved over four weeks anyway. And the field is genuinely split on whether to use it: Constant Therapy uses a real-time recognition engine to mark naming responses correct or incorrect, while Tactus Therapy deliberately uses self-scoring and record-and-compare instead.
Every one of those trials was in post-stroke aphasia. There is no equivalent trial in primary progressive aphasia, where the goal is holding on to function rather than recovering it.
The risk of a machine saying you said a word you did not
This is the part worth being blunt about. A verifier can be wrong in two directions, and they are not equally harmful.
A false reject is the machine failing to credit a word that was said correctly. If the app's response to uncertainty is to offer the next cue rather than display "wrong", the cost is one unnecessary hint. The person still ends the trial saying the word.
A false accept is the machine crediting an attempt that was not the word. The real cost is not the wrong tally; it is that a mispronunciation gets reinforced as correct and the person never hears the right form on that trial. There is a straightforward mitigation, drawn from how errorless learning works: whenever the app credits an attempt, it should also play the correct model. Then even a false accept is followed by the right version.
What a tool like this should never do on its own is tell someone they got it wrong, or show them an error count. The best-validated verifier in the literature agrees with human therapists about 89.5 percent of the time. That is good enough to decide whether to offer a hint. It is nowhere near good enough to tell a person with aphasia that they failed.
If you want to practise alone, here is the honest summary
- Do not expect a transcription app, a voice assistant, or a dictation feature to understand aphasic speech reliably. The published rates are 38.4 percent WER on connected speech and target-word detection in 41 to 45 percent of single-word naming trials for zero-shot Whisper.
- A tool that checks one known word at a time is attempting a much easier task, and the best published version of that reached 89.5 percent agreement with therapists, with 6.6 percent false accepts and 3.9 percent false rejects.
- No commercially available service has been validated on aphasic speech. Pronunciation assessment scores are unvalidated for this population and run low on slow, effortful speech.
- Any such tool should be calibrated on recordings of the specific person, with a partner labelling a set of attempts first, before anyone relies on it unsupervised.
- Keep the recordings. A person should be able to listen back, and a speech-language pathologist should be able to review what the machine decided.
- A tool that only ever credits or offers a hint, and never says no, fails safely. One that scores and marks errors does not.
Sources
- Sanguedolce et al. "When Whisper Listens to Aphasia," Interspeech 2024. ISCA archive
- Davudova et al. 2025. Application of Whisper in clinical practice: post-stroke speech assessment during a naming task. arXiv 2507.17326
- Evaluating ASR for aphasia: a framework for clinically relevant transcription performance. Aphasiology 2026. Taylor & Francis
- Koenecke et al. "Careless Whisper: Speech-to-Text Hallucination Harms," FAccT 2024. arXiv 2402.08021
- Comparative investigation of ASR platforms for aphasia assessment batteries. PMC9863375
- Zero-shot recognition of dysarthric speech using commercial ASR and multimodal LLMs, 2025. arXiv 2512.17474
- Barbera et al. 2021. NUVA: a naming utterance verifier for aphasia treatment. Computer Speech and Language. PMC8117974
- Latent Phrase Matching for dysarthric speech, Interspeech 2023. arXiv 2306.05446
- Google Research. Personalized ASR models from a large and diverse disordered speech dataset. Google Research
- Tailored ASR in global aphasia with dysarthria, single case. Frontiers in Rehabilitation Sciences 2026. Frontiers
- iTalkBetter phase II trial. eClinicalMedicine 2024. PMC11056404
- VoiceAdapt virtual randomised trial, 2025. PMC12440831
- Microsoft. Use pronunciation assessment. Azure docs, and characteristics and limitations
- Bridging neuroscience and AI in aphasia rehabilitation, 2026 review. arXiv 2603.22357