People With Aphasia Put Human Review Before AI-Generated Speech

When artificial intelligence sits between a person’s intention and their spoken words, fluency is not enough. The person needs a practical way to see what the system thinks they meant—and to refuse it when the reconstruction is wrong.

That principle emerged from R-SPEAK, an exploratory communication-support project for people with expressive aphasia. In its first design phase, three stroke survivors and one carer helped rank the features they considered essential. Their “must-have” list called for displaying a reconstructed phrase before use, reading it aloud, flagging incorrect output, and separating private from public conversations so private material would not shape later interactions.

The important distinction is that these were prioritized design requirements. The published paper describes the tested web prototype as recording a response and displaying one model-generated output during a supported holiday-booking exercise. It does not establish that every proposed control—especially the private/public boundary—was fully implemented and tested. Human agency was the design target, not yet a verified product capability.

Research references: [1] [2]

Co-design turned oversight into interface requirements

Aphasia can affect speaking, understanding, reading and writing after stroke or other acquired neurological conditions. A person may know what they want to communicate while finding it difficult to produce a complete, fluent sentence.

R-SPEAK combines automatic speech recognition, a language model and text-to-speech in a proposed pipeline. Speech becomes text; a model proposes a clearer reconstruction; the user reviews the result; and an accepted version can be spoken aloud. The architecture was designed around locally deployable open-weight models, with cloud access as an option.

The participatory work made “human oversight” concrete. The group did not simply request a general human-in-the-loop promise. It prioritized visible review, an audible rendering, an error flag and a privacy distinction. Facilitators used pictures, keywords, simplified graphics, extra response time and closed as well as open questions so communication difficulty did not exclude people from the design conversation.

That process is a useful model for assistive AI: the person affected by an error should help define where the system pauses, what it reveals and how it can be corrected.

Research references: [1]

The prototype evidence is promising and narrow

The team then evaluated a functional prototype with four stroke survivors with aphasia and one carer. Three of the four people with aphasia were rated as having moderate aphasia and one as mild. All had some experience with computers or smartphones, none had used AI, and none had used an augmentative-and-alternative-communication app, although one had used speech-therapy exercise apps.

Testing covered one supported scenario: booking a holiday through eight prepared questions. Participants used a shared web prototype during an interview, and a carer or speech and language therapist could attend to support communication. The four participants with aphasia gave the system a mean System Usability Scale score of 75, with individual scores from 62.5 to 85.

Those numbers describe perceived usability in a proof of concept. They do not show that R-SPEAK improved independent everyday communication, listener understanding, rehabilitation, quality of life or long-term participation. The authors themselves call the usability results exploratory because the sample was so small.

Clinicians were also more cautious than participants and carers. They worried that real conversation depends on context, implication and relationships that a prototype may not capture. Participants suggested better accuracy and speed, editable responses, rerecording, spoken output for people who find reading difficult, and multiple choices rather than a single interpretation.

Research references: [1] [2]

The model benchmark was not a clinical outcome

The technical development had two distinct stages.

First, the researchers selected 30 standalone question-and-answer pairs from AphasiaBank and tested eight open-weight language models. Five team members with clinical or technical expertise scored the outputs independently, then discussed differences. No formal interrater-reliability statistic was calculated. Mixtral 8x7B received the strongest aggregate evaluation and was integrated into the prototype.

Later, the team compared 12 smaller open-weight models on 30 clinician-patient dialogue samples. Each model generated ten interpretations per sample, producing 3,600 outputs. A larger Qwen3 8B model scored relevance, faithfulness and completeness, with a subset compared against human expert assessments.

Qwen2.5 3B ranked highest among the smaller models with a mean composite score of 0.814 and a mean response time of 0.600 seconds. Mixtral scored 0.838 in the comparison but averaged 2.42 seconds. That identifies a useful engineering tradeoff between output quality and conversational delay. It is not a measurement of clinical accuracy, safe autonomous use or improved communication in the world.

The benchmark’s reference interpretations were themselves expert reconstructions of isolated AphasiaBank utterances. The original speakers’ intended meanings could not be confirmed, and much of the later scoring relied on another language model. The paper also reports that Mixtral sometimes produced inaccurate or overly verbose interpretations.

Research references: [1] [2]

Accessibility and consent cannot be afterthoughts

The study’s consent process included aphasia-friendly pictorial information and an option to have the information sheet and consent form read aloud. Participants had to demonstrate informed consent. Sheffield Hallam University approved the study under ER64821246, and interview participants received £50.

Even with those accommodations, the tested group does not represent everyone who might need communication support. People with more severe aphasia, reading or auditory-comprehension difficulty, motor or visual impairment, cognitive difficulty, or limited technology experience may face different barriers. Cost and training may also determine whether a tool expands access or narrows it.

Research references: [1]

A proposal should remain a proposal

The most important idea in R-SPEAK is not that a language model can make fragmented speech sound fluent. A fluent mistake can still take authorship away from the person speaking.

Assistive AI should therefore make the boundary between proposal and consent visible:

• the model proposes an interpretation;
• the person can inspect and hear it;
• rejection and correction are easy;
• uncertainty is not hidden by polished language;
• private material stays private; and
• no output is treated as the person’s voice until the person accepts it.

What the prototype has—and has not—shown

R-SPEAK has not yet proved that this full pattern works independently in everyday life. It has done something earlier but valuable: it let people with aphasia help define what meaningful control should look like before a larger product and trial are built.

The article was published in the *Journal of Medical Internet Research* on October 8, 2026. The authors report funding from EPSRC project EP/W000679/1 and the NIHR HealthTech Research Centre for Rehabilitation, declare no conflicts of interest, and state that Qwen2.5 and Grammarly were used only for grammar and sentence-structure checks in manuscript preparation.

Research references: [1] [3]