Give your robot vacuum your own voice.
Replace the Xiaomi Robot Vacuum S20's voice prompts with your own recordings — entirely over your local network. No cloud account, no rooting, no soldering.
You get a small web app that runs on your own machine. It lists all 80 prompts, tells you when each one is spoken, lets you drop in a recording, plays it back through the robot's own speaker so you can check it before committing, matches its loudness to the factory pack, and installs.
On Windows: download, unzip, double-click Console.bat. The launcher
installs what is missing, creates the config file and opens it for you — you
only supply your vacuum's IP address and token.
Details below.
┌─ Voice console ──────────────────── Docked · 100% · volume 10/10 · fr_FR ──┐
│ 9 replaced of 80 [All] Replaced Original [Settings] [Install] │
├──────────────────────────────────────────────────────────────────────────────┤
│ Localisation · for finding it, and for it to find itself 5 prompts │
│ │
│ ▶ 11_1.mp3 " Je suis là, je suis là, je suis là " │
│ You press "find robot" in the app, to locate it under a sofa. │
│ 2.35 s +3.4 dB [On the robot] [▶ my voice] [Restore] │
└──────────────────────────────────────────────────────────────────────────────┘
The vacuum accepts a voice pack from any HTTP server — but only if the URL reproduces the path used by Xiaomi's CDN:
http://<your-pc>:8088/ijai-2s-map.record/multilanguage/xiaomi3/S20_20240110/fr_FR.tgz
Serve the identical archive from http://<your-pc>:8088/fr_FR.tgz and the
robot downloads it, reports 100 % progress — and silently ignores it. The
domain is never checked; only the shape of the path is. That one detail took
fifteen failed attempts to isolate, and it is why no DNS spoofing, no TLS
certificate and no cloud call are needed.
Three traps worth knowing before you experiment:
- Installing the language that is already active does nothing. You must switch to a different language code first. The app does this for you by bouncing through Xiaomi's official German pack.
- The status codes tell you nothing. State
15dropping to0within three seconds is the normal pattern — for success and failure alike. It is exactly what Xiaomi's own working URL produces. Only listening tells you. - Never test with a single replaced file. That mistake produced a wrong diagnosis for days: the pack was installing perfectly, but the prompt being triggered was a different file than assumed.
The full protocol, the measurements and every dead end are in FINDINGS.md.
Two double-clicks. The launcher does everything else.
-
Download. Green Code button above → Download ZIP → unzip it anywhere.
-
Double-click
Console.bat. It checks what you need, installs anything missing, then createsconfig.jsonand opens it for you. -
Fill in two values in the file it just opened, then save and close:
{ "ip": "192.168.1.42", "token": "0123456789abcdef0123456789abcdef" } -
Double-click
Console.batagain. It downloads the official voice pack and opens the app in your browser.
That's it. No terminal, no pip, no git.
| Where to get it | |
|---|---|
| IP address | your router's device list, or the Mi Home app under the vacuum's settings |
| miIO token | Xiaomi-cloud-tokens-extractor — sign in with your Xiaomi account, it prints the token |
The launcher handles the rest: it installs cryptography and
ffmpeg if they are missing, and fetches the official
pack. Python 3.9+ is the one thing it cannot install for itself — if it is
absent, the launcher tells you and points you at
python.org. Tick "Add python.exe to PATH"
during that install.
config.json is gitignored: your token never leaves your machine.
git clone https://github.com/LBougx/s20-voice.git
cd s20-voice
pip install cryptography # ffmpeg must be on your PATH
cp config.example.json config.json # then fill in ip + token
python telecharger_pack.py # fetch the official pack
python console.py # opens http://localhost:8088Set lang to your pack's language code and everything follows —
telecharger_pack.py fetches that pack and the app works on it:
{ "ip": "…", "token": "…", "lang": "de_DE", "interface": "en" }The interface picks its own language from lang, or you can pin it with
interface (fr and en ship today). If no prompt mapping exists for your
language yet, rows show filenames only and the app says so — the mechanism
still works, you just don't get the "what triggers this" column until someone
contributes one.
python-miiois deliberately not a dependency: itsnetifacesrequirement has no prebuilt wheel past Python 3.8. Everything here speaks the miIO protocol directly, withcryptographyas the only third-party package.
Each prompt gets a row: play the original, read what triggers it, drop in your own audio in any format — conversion is automatic — and hear it on the robot.
On the robot sends the sound to the vacuum and plays it through its speaker
in about five seconds. It borrows prompt 11_18, which nothing triggers in
normal use, so your other customisations stay untouched. Under the hood it calls
MIoT action 7.11 (set-listening-voice) — the sample the Mi Home app plays
when you pick a voice.
Worth being explicit about, because confusing them wastes time:
| Setting | What it changes | Takes effect |
|---|---|---|
| Vacuum volume (0–10) | the device's own setting, MIoT 4.2 |
instantly |
| Recording level (LUFS) | the gain applied to your files | after re-encoding |
Home recordings typically land 3 to 12 dB below Xiaomi's prompts — which is why your voice sounds like a whisper between two announcements. Eleven decibels is heard as roughly half as loud. The default target, −17.9 LUFS, is the median of the official pack.
Your source files are kept in mes_voix/_sources/, so changing the level
re-encodes from the originals — no cumulative loss however often you adjust it.
On by default. The 78 spoken prompts are 16 kHz / 32 kb/s, but the pack's two non-verbal sounds are 44.1 kHz — up to 320 kb/s, and one is stereo. The decoder handles far better than speech quality, so your recordings go in at 44.1 kHz / 128 kb/s instead of being crushed.
| File | Role |
|---|---|
console.py |
the whole tool: local web app, miIO client, pack builder |
Console.bat |
Windows launcher — checks Python, installs deps, opens the browser |
telecharger_pack.py |
downloads the official Xiaomi pack into packs/ |
lire_aspi.py |
read-only diagnostic; masks your token and MAC in its output |
data/fr_FR.json |
the 80 prompts: text, trigger, duration, grouped by situation |
data/ui.json |
interface strings, one block per language |
FINDINGS.md |
the protocol, the measurements, the dead ends |
Source comments are in French; the interface and this documentation are translated. The method applies to any language pack.
data/fr_FR.json may help you even if you never run the code: all 80
filenames mapped to what the robot says and when. That mapping was done by ear,
and the numbering is not sequential — 11_1 is the "find robot" phrase, 11_18
a shorter variant of it that nothing ever triggers.
The local protocol exposes much more, all without the cloud:
4.1— plays the locate prompt on demand, turning the robot into a remote-triggerable speaker for any script or automation7.16—1=Forward, 2=Left, 3=Right, 4=Back, 5=Stop: manual driving, which the Mi Home app does not offer on this model3.1,2.1,7.8–7.15— battery, status, and wear on brushes, filter and mop pad, for your own dashboards- services
8,9,10,12— schedules, virtual walls, maps, do-not-disturb
Full spec: miot-spec.org
So you don't spend time on these:
| Approach | Result |
|---|---|
| Public HTTPS hosting (catbox.moe) | the robot never downloads at all |
| Cloud commands | rejected with -704220043 before any download |
| Rebooting after install | no effect — the language label survives a power cut while the audio does not change |
| Rooting | xiaomi.vacuum.d106gl is on Dustbuilder's "very unlikely to ever be supported" list |
Two things are worth contributing, and neither requires touching the code.
A prompt mapping for your language. Identify the 80 prompts in de_DE,
en_US, it_IT… and drop a data/<lang>.json following the shape of
data/fr_FR.json. The app picks it up automatically. Expect it to take about
40 minutes by ear — the numbering gives no clue, and the app's built-in player
makes it bearable.
An interface translation. Copy the en block in data/ui.json under a new
two-letter key and translate the values. Nothing else to change.
python telecharger_pack.py --toutes sweeps 28 known language codes across the
three dated CDN folders, so you can check what exists for your region.
Reports from other firmware versions or nearby models are welcome too: open an
issue with the output of lire_aspi.py (it masks your token and MAC for you).
Everything here is reversible. If a pack misbehaves, change the language from the Mi Home app and the official pack comes straight back — that path was verified before anything else was attempted.
This talks only to a device on your own network, using your own token. Nothing is sent anywhere else.
Xiaomi's audio files are not redistributed here. telecharger_pack.py
fetches them on demand from Xiaomi's own public CDN.
MIT — see LICENSE. Xiaomi's audio files and trademarks belong to Xiaomi; this project is not affiliated with or endorsed by Xiaomi.