Speekl is open to everyone from today.
For the last stretch it has been in the hands of a great many of the people already here: active Speekl students, who agreed to practise on something half-built, and a large part of the tutor side of the marketplace, people who have spent years teaching others to speak and were happy to tell us, in some detail, where the AI was getting it wrong. That is the only reason this is worth releasing. A language tutor is not a thing you can specify in a document and then go and build; you find out what it has to do by watching a few thousand people try to use it.
So this post is less an announcement than an account of where the time went.
The model was the smallest decision
It is tempting to describe the last few months as wiring up a voice model. Honestly, choosing the model was among the quicker calls. We tested the providers against real learner speech: accented, hesitant, half in the wrong language, full of restarts. The differences that mattered were rarely the ones on the benchmark pages. What mattered was latency, and how the model behaved when the input was a mess.
The turn-taking budget
- Human conversation~200 ms
The gap people leave between turns. It carries meaning: hesitation reads as doubt, an instant answer reads as confidence.
- Speech-to-speech model~300–500 ms
Close enough that learners stop performing for the machine and start just talking.
- Speech-to-text → model → speech~2–3 s
Technically a conversation. In practice a walkie-talkie, and spontaneity dies inside five minutes.
Below roughly half a second, something changes in how people speak to it. They interrupt. They correct themselves mid-sentence. They stop waiting politely for their turn. That is the behaviour we were buying, and it is the reason latency was the first constraint we designed around rather than an optimisation we left for later.
Most of the work sits above the model
The model generates language. It does not decide what you should practise today, when to let a mistake go, when to stop the conversation and drill the thing you got wrong three times, or what to remember about you next week. Those decisions are the tutor. Everything we built is the machinery around the model that makes them.
What a lesson actually runs on
- Voice
Speech in, speech out, and the turn-taking in between: interruptions, pauses, and knowing when you have finished a thought rather than merely stopped talking.
- The harness
The part that behaves like a teacher: what this lesson is for, which correction is worth making now, when to push and when to let you finish the sentence badly and move on.
- Memory
What you keep getting wrong, what you are learning the language for, and what happened in the last six lessons, all carried forward so session seven is not session one again.
- The model
Generates the language. Swappable, and it has been swapped. It is the engine, not the car.
The harness
A general assistant will correct everything, because being helpful is what it was trained to do. A good teacher corrects almost nothing. They pick the one error that is actually blocking you and let the other six go, because a learner who is interrupted every sentence stops speaking in sentences.
Encoding that judgement took longer than everything else combined. When to interrupt. When to switch to your native language and when to refuse. How hard to push when you are visibly struggling, which is not the same as when you are merely slow. Every one of those turned out to be a rule we could only find by watching a tutor do it and asking why they had done it that way.
The memory layer
A tutor who does not remember you is a conversation partner. The value of a real teacher accrues over months: they know you keep dropping articles, that you are learning for a job in Berlin and not for a holiday, that you gave up on the past subjunctive twice already.
Getting that right is mostly a question of restraint: deciding what is worth keeping, what is worth resurfacing, and when raising something again is useful rather than nagging. Remembering everything is as useless as remembering nothing, and considerably more annoying.
The prompts
Prompts are not the product, and they still took months. Not writing them; writing one takes an afternoon. The time went on finding the cases where a sensible-looking instruction produces a tutor who is subtly wrong: too eager with praise, too quick to accept a broken sentence, too willing to answer a grammar question in English when the whole point of the session was to stay in the target language. You only ever find those by running real conversations and reading the transcripts, which is slow and unglamorous and has no shortcut.
How we got here
Sixteen months, two builds
May 2025
First prototype
A web app built to answer one question: can a model hold a language lesson at all? It could, badly, and it taught us what the product had to be.
December 2025
We stopped adding to it
The prototype had answered its question and was the wrong shape for the answer. Rather than keep extending it, we froze it.
June 2026
Rebuilt from scratch
The current app, designed around latency, the harness and memory from the first commit instead of having them bolted on.
September 1, 2026
Public release
Open to everyone, in every language we support.
Where we are on day one
Speekl today
- learners practising on Speekl
- 20,000+
- minutes of conversation spoken
- 1.4M+
- languages and dialects
- 25+
- average learner rating
- 4.8/5
This is the start of it
The list of things we know are not good enough is longer than the list of things we shipped. Memory should reach further back. The plan should adapt faster when you obviously do not need the lesson it picked. Some languages are noticeably stronger than others, and we would rather say so than let you find out in session three.
What is different from a month ago is that we are no longer guessing about any of it. Every one of those came from somebody using the thing.
If you have been waiting for an invite, there is no longer anything to wait for. Start a conversation, get something wrong, and let the tutor tell you about it.