Helpful support, clear limits, human judgement.
EaseTalk is used by people working hard to be understood, and by the clinicians treating them. This page sets out how we build AI for that, including the part most products leave out: speech recognition mishears atypical speech, and pretending otherwise would be the easiest way to do real harm here.
Helpful support, clear limits, human judgement
That sentence is the whole of our position on AI, and the rest of this page is what it means in practice.
The people who use EaseTalk are often working hard to be understood. Building AI for that audience carries an obligation not to overstate what it can do, because an overstatement lands on somebody who is already being told, often, that the problem is them.
So: a speech and language therapist stays in charge. Anything the AI produces can be seen, corrected and ignored. We collect what the service needs and not more. And we are specific about misrecognition, because a system that mishears atypical speech and then quietly counts it as a mistake would be the single worst thing we could build.
1Clinical oversight
EaseTalk is built around the relationship between a person and their speech and language therapist. It is not built around the software, and the software is not trying to become the therapist.
In practice that means the therapist assigns the practice, the therapist reviews what came back, and the therapist decides what any of it means. The app has no mechanism for changing a plan of care, flagging a clinical concern as fact, or acting on anything by itself. If EaseTalk were removed tomorrow, the clinical relationship would still be intact, which is roughly the test we hold ourselves to.
We also do not build features that quietly move clinical judgement into the product. If a feature would only make sense on the assumption that the AI knows better than the clinician, that is a reason not to build it.
2Reviewability
An AI output nobody can see is an AI output nobody can correct.
- Visible. The words the speech recogniser detected are put on screen during practice, under What we heard. Summaries shown to a therapist are text they read, not a score they are handed.
- Correctable. Your reflections are your own words and are not rewritten by the system. A therapist can edit or replace anything the AI drafted before it becomes part of their own notes.
- Disregardable. A therapist can ignore an AI output entirely, and nothing in the product treats that as an error, an override, or something to be justified.
3Data minimisation
We collect what is needed to run the service and to support clinical review. The clearest example is audio.
No recording is kept of your practice
Your microphone is streamed during a live practice so your speech can be recognised, and then the audio of your turn is cleared rather than filed against the session. Sound Setup measures the noise in the room and your voice level, keeps those measurements on your own device, and saves no recording. Storing your voice would be easy and it would be useful to us. We do not do it.
Equally, there is no analytics toolkit in the app, no crash reporting toolkit, no advertising, and no third party tracking library. This website sets no cookies and loads nothing from anyone else's servers. Signing in sets one cookie, which holds the tokens that keep you signed in and does nothing else. Those are not absences we forgot to fill in. They are the design.
Where we do send data to AI providers, that is set out precisely in the privacy policy rather than described as necessary technical processing.
4User control
You decide what leaves your practice.
- A therapist who assigned you practice can always see whether you did it, because the assignment is theirs. That is the limit of what is automatic.
- Everything about how the practice went, the objectives, the detail, the trends and your reflections, sits behind a separate, optional consent. Refusing it does not stop you using EaseTalk.
- Either way, your therapist does not see the full transcript of your conversations. Practice needs somewhere it is safe to be bad at something.
- You can end any practice at the point you choose, using end practice. Nothing is scored against you for stopping.
- You can withdraw a consent, ask for your data, or ask us to delete it. The route that works today is described in the privacy policy.
One honesty point on that last one. The in-app control for withdrawing therapist sharing is not working yet and shows a coming soon message, and the app's export and delete buttons were removed in August 2026 because they did not work either. Rather than leave buttons that pretend, we point you at the route that does work: emailing us. We would rather be embarrassed on this page than have you believe you have withdrawn something you have not.
5Fairness and misrecognition
This is the section that matters most, and it is the one you should judge us on.
The problem, stated without softening
Speech recognition systems are trained mostly on typical speech. They perform measurably worse on atypical speech. Dysarthria, stuttering and aphasia are exactly the kinds of speech they handle least well.
The people most likely to be misheard are the people we built this for
That is not an edge case in EaseTalk. It is the central population. Any speech product for this audience that does not say so out loud is either not paying attention or hoping you are not.
What we do about it
- We show you the transcript. The words the recogniser detected appear on screen as What we heard. A mishearing is visible, not silent. This is the single most important design decision in the product.
- We calibrate. Sound Setup measures the room noise and your comfortable voice level so the recogniser is working in better conditions.
- We keep a clinician in the loop. A therapist reading a reflection can tell the difference between a hard conversation and a bad transcript. Software usually cannot.
- We treat correction as information. When the recogniser is wrong, that is worth telling your therapist. It is data about the tool, not a failure by you.
What we will not do
- We will not silently score somebody on a recogniser's errors.
- We will not present a recognition confidence number as if it described the person's speech.
- We will not build a feature that ranks users against each other on anything a recogniser produced.
- We will not describe accuracy in marketing without saying who it is accurate for.
What we have not done yet
We have not published our own measurement of how much worse recognition performs on atypical speech within EaseTalk specifically. We know the effect exists. We have not quantified it in our own product, and we are not going to imply that we have. It is on the list of things we have not shown.
6Safety and escalation
EaseTalk is not a crisis service
Nobody monitors the app in real time for emergencies, and it must not be relied on to raise an alarm. If you are in distress or need urgent help, contact your treating clinician or your GP, or call the emergency services on 112 or 999.
Practice conversations can touch on difficult things, because the situations people need to rehearse are often the ones that frighten them. Our servers do store safety information relating to a session, and where you have consented to share with a linked therapist, a clinician reviewing your practice is the person positioned to notice something and act on it.
We are not going to describe the mechanics of that in more detail than we can stand over. What is written above is what we can state accurately today, and it is deliberately modest: the person who notices is a clinician, not the software.
7Transparency
The plainest way to be transparent about an AI system is to be specific about what it does not do. So, in one place:
It does not diagnose
No condition is identified, suggested or ruled out.
It does not assess or measure clinically
Confidence and difficulty patterns describe what you logged about engagement. They are not an outcome measure and have not been validated as one.
It does not decide or recommend treatment
No plan of care is created or changed by the software.
It does not act on its own
There is no automated decision with legal or similarly significant effects on you.
It does not pretend to be a person
You are told you are practising with AI characters, and the app does not blur that.
It does not learn from you to improve itself
Your identifiable data is not used to train models. See the AI terms for the current status of that commitment.
The providers behind these features are named in the AI terms and in the privacy policy, rather than being described as trusted partners.
8Holding us to this
A page of principles is worth very little on its own. Two things make it checkable.
First, specifics. Our privacy policy names every company that touches your data and says which country it is processed in, rather than describing them as trusted partners. If a claim on these pages is ever too vague to check, that is a fair thing to challenge us on.
Second, tell us when we get it wrong. If the app produced something that misdescribed you, if the recogniser made a mess of your speech and something downstream treated it as fact, or if anything here does not match your experience, write to easetalkai@gmail.com. Those reports are how the fairness section above stops being a statement and starts being a practice.