Deutsch · English
Type a sentence, get an audio file back — and they all sound alike.
I am building this for my three-and-a-half-year-old daughter, who uses augmentative communication.
mitreden is German for joining in a conversation.
People who use augmentative communication often speak through more than one device: a talker app on a tablet, a reading pen, single buttons. Each brings its own voice along. The same sentence then sounds like a different person depending on the device — and someone who is only just learning that these devices are their own voice notices immediately.
mitreden turns that around: the voice is chosen once, the sentences live in one place, and every device gets the same audio file. Change the voice later and one command makes everything match again.
It is not a talker and does not replace one. It is the workshop behind it: text in, audio files out, in the format your device understands.
The sentences are the source of truth — a list, kept in the browser. The audio is the result, one file per sentence, made on the spot and downloaded when you want it.
Because the audio can be made again at any time, the sentences are the only
thing that really needs a backup. They come out as phrases.json, a plain
list you can read.
The repository is deliberately small: a page, the recording chain it runs, and the words it says in two languages. No framework, no build step for the interface, and no dependency it fetches while you are using it. The code is in English, the interface speaks German and English, and this guide exists in both languages. Commit messages are English too — the history is a developer document, and the sibling projects settled on the same rule.
A browser. Nothing else.
Open mitreden and start typing. There is no account, no key, no installation, and nothing to keep running. The voice is downloaded once — about 60 MB the first time, then it stays on your machine — and after that every recording happens in the tab. Nothing you type is ever sent anywhere.
If you used mitreden before and it fetches a voice you already had, that is expected and not a failed download. The page now runs piper itself instead of handing the job to a library — which is what makes Kerstin and John speakable at all — and it keeps its models in a store of its own. It happens once per voice, and then never again.
Type sentences into the box, one per line, and press "Add sentence". Each one is spoken and appears in the list below with a player. Tick the ones you want and download them as MP3 or WAV, then copy them onto the talker, the reading pen or wherever they need to go.
The sentences stay in the browser you typed them in. That is the one thing worth a copy: Settings → Save sentences to a file. Clearing the browser's data takes them with it, and unlike the audio they cannot be made again.
For more voices, see Choosing a voice.
The language picker sits at the top right. On your first visit mitreden
follows your browser — German for a German one (de, de-AT, de-CH), English for
everyone else — after that whatever you chose last; the choice is also in the
address (?lang=en), so a link carries it along.
The language of the interface and the language of the voices have nothing to do with each other: both versions offer the same voices, German and English alike.
The texts live in lang/de.json and lang/en.json, with English keys.
Another language is another file — it shows up in the picker without a code
change. A missing key falls back to English, and if it is missing there too,
the key itself appears: a gap should be visible, not blank.
One file per sentence, named after its id. The id stays out of the list — it is a file name, not something to read. It appears in one place, when you change a text, because that is where it matters: the file name stays as it is, and what is already on a device keeps working. It is built from the first words of the sentence and stops at a word boundary, so a long sentence does not become a long file name. Two sentences with the same beginning get a number appended. Existing ids stay as they are — they may already be written on a device.
Recordings are MP3 at 44.1 kHz mono, which just about every device and app understands. WAV comes out of the same menu when a device insists on it — converted on the way out, from audio that was already trimmed and levelled, so the two sound the same.
Every file is silence-trimmed and normalised to −16 LUFS. Without that one sentence is barely audible and the next one shouts.
The voice sits next to "Add sentence". It decides what recordings are made with: every new sentence gets it, existing ones keep theirs until you tick them and pick "Change voice". One voice for everything stays the normal case — "Select all", record once, done. But when a sentence belongs to someone else, or is in another language, it may sound different.
Every row says which voice it was recorded in. A voice is called
Kerstin · piper · de: name, where it comes from, which language it speaks.
The language is there because that is the reason two voices in the same list
really differ — in a household with two languages, say.
To move sentences already recorded, tick them and choose "Change voice" from the actions menu. "Select all" first makes every sentence match.
Cloud voices — Azure, ElevenLabs — are deliberately not offered. A key typed into a public web page is a different promise from the one this makes, and it is not one worth making for a handful of extra voices: the four here are free, offline after the first download, and cannot be switched off by a company changing its mind.
piper is what speaks, compiled to WebAssembly and running in the tab. The models come from the piper project's own publication on Hugging Face, fetched once and then kept on your machine. It will still work in ten years, without an account and without a subscription, because nothing about it depends on a service staying up.
Past a few dozen sentences you want some order. mitreden uses collections, not folders — one sentence can be in several at once.
That is not a detail of the interface, it is the point. A collection does not own its sentences; it points at them. If it owned them, a sentence you use at nursery and at home would be two sentences, and two sentences are two recordings that can drift into two different voices. One text stays one sentence and one audio file, however many collections it belongs to.
{ "id": "i-need-help", "text": "I need help.",
"collections": ["nursery", "home", "emergency"] }Above the list are two rows to click: Collections and Voices, plus a search box. The voices row also holds "Not recorded", in case something is missing. Several can be picked at once, and the rows work together: collection "play" and voice "Kerstin" shows what is both.
How to tell filters from actions: filters are pills, actions are boxes. The pills only change what you see. The buttons in the separate box below do something — and they only appear once you have ticked some sentences. What collections a new sentence joins is decided solely by its own field at the top.
Several collections can be picked at once and combine with OR; the free text
narrows it further. The search runs in the browser, without waiting, over text
and collection names, and is forgiving about umlauts: hor auf, hoer auf and
Hör auf all find the same sentence.
The row shows the twelve most used collections and folds the rest behind "+ n more". Everything else about a sentence sits in the ⋮ at the right edge of its row.
One text, one file. Add a sentence that already exists and no second entry appears — the existing one just picks up the new collection. Capitalisation and extra spaces do not matter, punctuation does: "Again!" and "Again." are spoken differently, so they are two sentences.
Loading a file works the same way: sentences that are already there are merged rather than duplicated, and you are told how many were.
Selecting. Recording happens by itself — when you add a sentence and after you change its text. The checkboxes in front of the rows are for the two things that concern several sentences: downloading and switching to another voice. Both buttons appear as soon as something is ticked: "Download as MP3" directly, everything rarer behind the chevron next to it — as WAV, change voice, add to a collection, remove from one, delete. It is the same menu the ⋮ of a row shows for a single sentence.
For several sentences there is adding and removing, but no replacing. Ticked sentences usually sit in different collections, and a filter may hide some — replacing would silently throw away something you cannot see. In a single row "Change collections" still replaces, because there the existing ones are in the field.
Should a recording fail, the "Not recorded" pill in the voices row says so. Click it, select, set a voice — and they are there. A filter leaves the selection untouched and tells you when something selected is currently out of sight. A single sentence can also be recorded again from the ⋮.
Downloading. The button packs up the ticked sentences — as MP3 directly, as WAV through the chevron next to it. The format is independent of how the files were recorded; that is for the one device that does not fit in. The conversion happens on the way out — what is stored stays as it is.
A single file comes out of the ⋮ of its row, in the same formats.
Nothing you type leaves your machine. There is no server to send it to: the sentences live in the browser's own storage, the voice is downloaded once and then works offline, and the audio is made in the tab. The only request that ever goes out is the one that fetches a voice model, from the piper project's publication on Hugging Face.
That has one consequence worth taking seriously. The sentences are in that browser and nowhere else. Clearing its site data deletes them, as does a browser that decides to reclaim storage. The audio can always be made again; the sentences cannot.
So: Settings → Save sentences to a file, and keep it somewhere real. What comes out is a plain JSON file you can read, and loading it back merges rather than overwrites, so a backup can never cost you work that came after it.
The interface is index.html plus the modules under src/; src/i18n/*.json
is every word it says. Vite builds it, TypeScript checks it, and both run off
plain npm — the same setup as
bildhaft, on purpose.
The speaking and the recording chain are not here. They come from stimmquelle, which vorlaut uses too — two products that both put a sentence on a child's talker cannot afford to disagree about how loud it comes out. It is installed from npm, pinned to a commit.
git clone https://github.com/Lautstark/mitreden.git
cd mitreden
npm install
npm run dev
Nothing is fetched from a CDN at runtime: the wasm the voice runs on is copied
out of node_modules into the build, and a missing file fails the build rather
than quietly reaching for a third party. The one request the running page makes
is the voice model itself, once, from where the piper project publishes it.
This is a hobby project, grown out of a concrete need at home — there is no roadmap and no promises. If you use it for your own child, or for someone you support: good. Bug reports and questions are welcome, even if answers may take a while.
If you change something: npm run test:e2e builds the site and drives the
real production bundle in a browser — Playwright, installed with
npx playwright install chromium. npm run typecheck is the fast half and
catches the mistakes that are actually made while moving code. Both run in CI
on every push.
The site is published by that same workflow rather than straight from the branch, and only once everything is green, so a commit that breaks the page reaches nobody. It needs Settings → Pages → Source: GitHub Actions.
If you are building something similar and could use advice, open an issue. There is little tooling for augmentative communication that you can hold in your own hands, and the experience of it travels badly by other routes.
bildhaft is a companion project: type a German sentence, get a row of AAC pictograms back, correct them, print them. It runs entirely in the browser too. The two are the same sentence seen twice — bildhaft gives it symbols, mitreden gives it a voice — and mitreden's importer reads bildhaft's collection files, so a picture book translated over there does not have to be typed again over here.
They share a mark: the same speech bubble, drawn from the same path, in a different colour. And a principle. bildhaft stores symbol references rather than pictures; mitreden stores sentences rather than audio. Either way the exported file still works for the person who receives it, whichever symbols or voices they happen to have.
MIT — see LICENSE. Do what you like with it.