Talking instead of typing
Speaking works the same way everywhere.
How it works
Every conversation — the main chat, a thread, a project's Creation Console, a viewer's chat on a shared work — takes a voice message the same way: the microphone sits to the left of the typing box, in all of them. In a role-play practice there is also 🗣 conversation mode beside it (below): turned on, replies are read aloud and you answer without pressing anything at all, and the typing box and the microphone give way to one bar. No other conversation has it for now.
- Press and hold the microphone button and speak. While you hold it, the typing box steps aside: it goes quiet, turns red at the edge, and says hold to speak… release when done. Bars beside the button move while it is listening — if they stay flat, your microphone is muted or the wrong one is selected. Your finger can slide off the button while you speak; it keeps recording until you lift it.
- Let go when you are done. Nothing is sent yet: your recording stays on your device, in place of the typing box, with a player. Play it back. If it is not what you meant, the ✕ drops it and the box comes back; holding the microphone again records over it.
- Send. Only now does the recording go to kuyun. The recording is the message: what kuyun heard is written in the conversation, marked transcribed from a voice message, with a round ▶ button beside that mark that plays your recording (⏸ pauses it) — so you always know why words you never typed are there, and you can hear yourself again later, on any device you open the conversation on. Playing one stops any other that is playing.
The reply comes back the same as for a typed message.
A quick tap records nothing. Letting go always ends the recording, so you cannot walk away with the microphone still open; dragging off the button ends it too.
In a course, that is how you say your answer out loud, hear yourself again, and see how it was understood, all in the thread.
Kuyun never hears the recording — it reads the written-down words. So if the transcript is wrong, the reply answers the words, not what you said.
If kuyun could not make out any words, it says so and keeps your recording, so you can play it back and try again rather than start over.
If the recording could not be turned into text at all, Kuyun says so in one line and asks you to try again, holding the button until you have finished speaking. The cause is on our side or the transcriber's — a recording too short to hold a word, a service that is down — never something in what you said.
Your recordings are kept with the conversation they are in. A project's go when you delete the project. Your main chat's stay until you delete all your data or your account, which destroys every recording you have made — a single conversation cannot be deleted on its own. There is more on deletion in Your data.
Languages
It works out which language you are speaking. To keep a single word from being misheard as another language, kuyun tells the transcriber what to expect: first the language you are practising — the one the last few replies in that conversation were written in, or the one an exercise is in — then the language your app is set to, and English. So a short Japanese answer in a Japanese role-play is expected to be Japanese even when your app is in English. Mixing those in one sentence is fine. If you speak a language that is none of them, it usually still gets it — but a very short clip may come back in one of the expected ones. A brand-new conversation has no replies to go by yet, so its first recording is expected in your app's language or English.
Hearing text read aloud
The other direction works too. The 🔊 and 译 (translate) buttons under a message in a language you do not read are switched off for now, except in a role-play practice, where the partner's lines always have 译 and a ▶ that plays the line again (A project's two conversations); the people running Kuyun may turn them on again everywhere. In a course, the text you are meant to hear has its own 🔊 button. There is more on this, and on translation, in Studying with a project.
The voices are OpenAI's. When OpenAI cannot be reached, Google's Gemini reads instead, and after it Alibaba's Qwen — so on a rare day the voices sound different. A conversation is always read whole by one of them, never half in one set of voices and half in another. The same goes for your recordings: OpenAI writes them down, and Google's Gemini when it cannot. Which companies, and what they receive, is in The AI models behind Kuyun and Terms of Service and Privacy Policy.
Listening to a book
A book or a course can carry a narration of each chapter, made by the people who made the work: a sound file beside the chapter, named like it. Such a chapter shows 🎧 in the project's files, and its page opens with a 🎧 Listen button above the text — and the same button at the top of the reader (Studying with a project, "Reading a work"), so you can listen while you read. Pressing it starts the player, a bar at the foot of the app with play, pause, seek, the listening speed and the chapter's length. The bar stays while you move about — to the Shelf, to a chat, to another work — and when a chapter ends the next one with a narration follows by itself, with your phone's screen off too; ⏮ and ⏭ move between chapters, and 📖 opens the chapter you are hearing in the reader, at the place where you stopped reading it — the same chapter, not the same sentence, since the narration does not know which sentence it is at. A sleep timer in the bar stops the reading after 15, 30 or 60 minutes, or at the end of the chapter. While a chapter plays, your phone's lock screen, a headset's buttons and a keyboard's media keys show the book and the chapter, and their next and previous move by a chapter, wherever your browser passes them on. Kuyun keeps where you stopped with your account, so you can start a chapter on your computer and continue on your phone; a chapter heard to the end starts over next time, and your Shelf card says where you are ("chapter-3 · 12 min in"). If your connection drops while it plays — a phone moving between networks — the player tries again by itself and picks up where it was. Your reading companion knows which chapters have a narration and where you stopped, and can take up the text from there — it does not play the sound itself; the player does.
If you made the work, Kuyun can do the reading. Narrate… in the project's ⚙ menu lists the chapters: pick one, some or all, choose a woman's or a man's voice for the whole book, and ask; a chapter's own page has a 🎧 Narrate button for just that one. Kuyun reads them in the background, usually a few chapters every few minutes, and each recording appears beside its chapter as it is done. You pay once, in tokens, as each chapter is made — about one token a character, so a 20-minute chapter is around 18,000 tokens. Your readers' listening counts against their own day, as reading a chapter does (Tokens). If a chapter cannot be read that day, the list says why, and asking again queues it once more. A recording stays until you ask for a new one.
Slowing it down
Spoken audio can play slower, which helps when you are learning the language. Beside every player there is a small 1× button: tap it to step through 1×, 0.85×, 0.7× and 0.5×, then back to 1×. The same choice is under Listening speed on your Account page. It is saved to your account, so your phone and your computer use the same speed, and it applies everywhere kuyun reads to you: a message, a course chapter, and the sentences in an exercise. The voice keeps its pitch when slowed. Your own voice messages always play at the speed you said them, and so does a paper's recording that plays once, while you are sitting the paper.
If it is not working
- The first time, your browser asks permission to use the microphone. If you said no, the button will not work until you change it in your browser's site settings.
- If the microphone button is missing, the browser does not support recording. Chrome, Safari and Edge all do.
- If you get a message about transcription not being available, that is a configuration problem on our side — tell the operator rather than retrying.
Conversation mode
Conversation mode is in a role-play practice only, for now: its chat has 🗣, just left of the microphone, which turns the mode on and off. No other conversation — your main chat, a thread, a project's Creation Console, a viewer's chat on another kind of work — shows it, and nothing there is read aloud or listened for by itself — except a mock coding interview, whose interviewer has 🗣 beside its box once the interview starts. The people running Kuyun may bring it to every conversation later. It is off until you turn it on. In a role-play practice, ▶ Start the first scenario turns it on for you, so the partner speaks first and the microphone is waiting when it finishes — for that visit, even if you turned 🗣 off there before: starting a scene is asking to talk. So do the chips that start the next one — ▶ Continue, ▶ Next and 🔁 Play it again. Your own choice stays as it was for your next visit. A scene started by a link from somewhere else (a collection's page, another tab) turns it on only if you have not turned it off, and a live conversation then waits for your tap on the bar. Whatever you choose with 🗣 is remembered on that device until you sign out. Someone else signing in on the same device starts with it off.
Turned on, it is a conversation rather than messages, and the typing box, the microphone button and Send give way to one bar: there is nothing to hold and nothing to type while it is on. What you had typed is kept, and comes back when you turn 🗣 off, which is also how to type or paste something. The bar is a wave that moves differently for each moment of the conversation, and says what it is in words beside it:
| The bar | What is happening | A tap on it |
|---|---|---|
| a rolling wave, green — Kuyun is speaking | a reply is being read aloud | stops the voice and opens the microphone, so you can answer straight away |
| your own voice, red — Recording you | the microphone is open and hears you | ends your turn and sends it at once |
| a slow breath — Your turn | the microphone is open, waiting for you to speak | ends your turn and sends what it recorded |
| a pulse moving along a grey line — Kuyun is thinking… | your turn was sent and the reply is on its way | nothing: it is coming |
| still — Tap to talk | the microphone has stopped listening | opens the microphone again |
If your device is set to reduce motion, the wave stays still and the words on the bar say the moment.
- Each reply is read aloud as it arrives — the replies to what you say or send. In a role-play, the partner is heard in a voice that fits the person it plays: the barista the scenario describes as a young woman in a woman's voice, Kevin at the bank in a man's — and the ▶ under a reply plays it again in that same voice, from the same recording. The voice starts as soon as its first words are ready, while the rest is still being prepared; in a browser that cannot play it that way, it starts once the whole reply is ready. Turning it on reads the newest reply on the screen first, unless it is a daily digest, which is never read aloud. Nothing is read while the page is out of sight, and the microphone does not open there either. Your browser plays sound only after you have tapped something on the page, so the first reply after you open it is read once you have tapped anything — a starting button, Send, the microphone.
- Then the microphone opens by itself. The bar says Your turn, and once you speak it turns red and moves with your voice, as the bars do when you hold the microphone. Speak, then pause. Kuyun listens to HOW you stopped, to tell a finished answer from a breath in the middle of one: when you sound finished, what you said is sent after about half a second of quiet; when you sound as if you are still thinking — "so first I would, um…" — it waits for you, for up to three seconds of quiet, and sends it then. This listening happens on your own device and the sound goes nowhere for it. It is a judgement, and it can be wrong both ways: it may wait a moment longer than you needed, or take a long breath for the end. The first time on a device your browser fetches what it needs for this, about 12 MB, and keeps it; until that has arrived, and in a browser that cannot run it, a pause of a second and a half sends what you said, however you stopped. Or tap the bar when you have finished, which sends it at once, even if you spoke too quietly for the pause to notice. No preview and no Send.
- If you say nothing for about ten seconds, it stops listening and the bar says Tap to talk: tap it when you are ready and the microphone opens again. Conversation mode stays on. To type instead, or to record a message and play it back before you send it, turn 🗣 off.
- To answer straight away, tap the bar while a reply is being read: the voice stops and the microphone opens.
- The screen stays on. A conversation needs no taps, so while conversation mode is on your phone or tablet does not dim and lock in the middle of it. Turn the mode off, or leave the page, and the screen sleeps as it usually does. A few browsers cannot keep the screen on, and some phones do not allow it while their battery saver is on; there it sleeps as before.
- Answers come sooner, with less thought behind them. While conversation mode is on, Kuyun is asked to answer quickly, as a person talking does, so it weighs a reply less than it would a message you read. For a question that needs care, turn the mode off and ask it there.
A study partner's conversation is live
With a work's study partner, conversation mode is one open conversation rather than turns: tap the bar once and just speak. The bar says Live — just speak; the partner hears you as you talk and answers in its own voice, and you can speak over it to cut in, or tap the bar to stop its voice. (When OpenAI's model holds the conversation, only the tap cuts in: it does not listen while it is speaking, so that it never mistakes its own voice for yours.) Nothing is recorded, sent or waited for, so the bullets above about ending a turn do not apply while it is live.
Started with one of those chips, the scene opens as written — its card, its picture, the partner's first line on the screen — and then the conversation opens by itself and the partner says that first line aloud, so the scene starts in sound. You answer by just speaking.
- What each of you said appears as text in the conversation once an exchange is finished — your line and the partner's — and stays there. Your line is text only: there is no recording of it to play back.
- It is spoken only. A live conversation shows no scene card, feedback card, picture or quick replies, and marks nothing finished. For those, turn 🗣 off and write, or record a message: the partner answers in full there, in the same conversation.
- The partner has one voice, whoever it plays.
- It ends after about fifteen minutes, or when you leave the page or switch to another app. The bar says Tap to talk; a tap starts it again and it goes on from the conversation's last lines.
- It counts against your day as a written turn does, by what was heard and said; with today's allowance used up it cannot be started.
- If it cannot be opened — the service that holds it does not answer, or your network does not reach it — conversation mode works the recorded way described above instead, from your next tap, without asking you.
- Your voice goes to the company whose model holds the conversation (The AI models behind Kuyun says which; the terms name it), for as long as it is open: straight from your browser when it is Google's, and through Kuyun when it is OpenAI's. Either way Kuyun keeps the words, not the sound.
In every other conversation, and whenever a live one cannot be had:
What you said is still a voice message in the conversation, with its transcript and its ▶ button, exactly as above. What is read aloud is the reply's words: in a role-play, the scene card and its line, the feedback card and a picture stay on the screen and are not spoken, and a line the partner gives you to hear (a listen block) is. Nothing is recorded while a reply is being read. A pause in which nothing was heard sends nothing; a tap on the bar while it listens sends what it recorded, and if there are no words in it Kuyun says it could not make any out. To stop listening without sending anything, tap 🗣, which turns the mode off. Leaving the page closes the microphone (A project's two conversations).
How a turn ended is counted, and nobody is named. With each turn of a conversation, Kuyun adds one to a count of how turns ended that day — it heard that you had finished, the quiet ran out, you tapped the microphone, or you spoke for a full minute — in the language you spoke. Not who you are, not what you said, not the time, and no sound: only totals for the day are written down, never a turn by itself, and they are kept. The team reads them to see, language by language, whether Kuyun is waiting too long for people to finish.