Blog · 5 September 2026
Why Parlessa has no speech recognition
Every speaking app scores you with a microphone. We chose not to, on purpose. The reasons are about learning, not about technology.
The first question almost everyone asks is “how does it know if I said it right?” The answer is that it does not know, and that we decided it should not try. This post explains the decision, because it is the one that makes Parlessa look old-fashioned next to the apps that listen to you, and because we think it is the most important product decision we made.
What speech recognition does in a language app
In the apps that use it, the microphone does one of two jobs. Either it gates you (you cannot move on until the recogniser accepts your sentence) or it scores you (a percentage, a colour, a number of stars). Both feel like feedback. Neither is the feedback a learner needs.
The recogniser is trained to guess what a speaker meant from what they said. That is the right goal for dictation and the wrong goal for pronunciation. A model that is good at its job will accept je les ai acheté for je les ai achetés every time, because in speech the two are identical; it will accept j’ai faite for je l’ai faite because it is optimised to be forgiving. Users of every major app say the same two things in their reviews: it passes anything, and when it does fail, it fails on sentences they said correctly. Both complaints are true, and they are two sides of one fact. The recogniser is not measuring what you want measured.
What you want measured at B1 is not “did the machine understand me” but “did I produce the right form, on time, with the right sound”. Agreement on the participle. The liaison in je les ai. The y in the right slot. The difference between pars and part, which is nothing in the ear and everything on paper. A recogniser cannot hear most of that, and a score that ignores it teaches you that it does not matter.
What replaces it
The lab’s answer is older and simpler: the correct answer is played to you two seconds after you say yours, and you compare.
You say Oui, je l’ai fait hier. The voice says Oui, je l’ai faite hier. You hear the difference (a consonant you did not say) while your own version is still in your ear. Then you say it again in the repeat pause, matching the voice. That comparison is the exercise. It uses the one instrument that is reliably better than any speech model at judging French: a human ear that has just heard a native speaker say the sentence.
It has other properties that a score does not:
- It is immediate. No upload, no spinner, no round trip. The lab keeps its rhythm.
- It is honest. The recording is right. Your ear will sometimes miss the difference, and that is fine: the structure comes back in a later lab and you get another pass at it. Over ten exposures you will hear it.
- It cannot be gamed. There is no threshold to sneak past, so there is no reason to mumble the hard part.
- It works with no screen. A lab runs with the phone in your pocket, on a walk, in a car. A microphone app needs you to hold the phone up and wait for the tick.
The reasons that are not about learning
There are three more, and we would rather say them out loud.
Privacy. Speech recognition means audio leaves your device, or a model runs on it. We did not want to write a privacy policy that has a paragraph about what happens to recordings of your voice. Parlessa has no microphone permission. There is nothing to leak.
Where you practise. People do not speak French out loud in a shared flat or an open-plan office if a phone is judging them. They do if nobody is listening. Serious learners tell us this without prompting: the reason they stopped using the app with the microphone was that they only felt able to use it alone at home, and they were never alone at home.
Cost and speed. Every recognised utterance costs money and adds latency. Removing it lets a lab load in a second on a bad connection, play offline, and cost us nothing per session, which is why the first unit can be free for good.
What we lose
Honesty requires the other column. Without a microphone we cannot tell you that you skipped an item, or that you said nothing for the whole lab. We cannot give you a score to watch go up. We cannot flag the one sound you always get wrong; you have to notice it. Some learners want the number, and for them a scored app is a better fit. We would rather be the tool that makes the number unnecessary than the tool that fakes it.
There is one thing we may add later: local self-recording, so you can hear your own attempt back next to the model, on your device, never uploaded. That is a mirror, not a judge, and it fits the method. A judge does not.
The short version
The lab does not need to hear you. It needs you to hear the difference between what you said and what a French speaker says, immediately, and to say it again. That has worked since 1958. The microphone is a solution to a problem the lab never had.