🎙️ Sado

One voice. Every language. Everywhere you are.

Sado is a voice-first AI companion that hears you, remembers you, and gets things done —
in the language you actually speak, not the one AI happens to support.

Built on Meta's Muse models for the Meta Global AI Developer Hackathon 2026 · Indie track

---

💡 The problem

Today's AI assistants are brilliant — in about 25 languages.

Meta's own real-time speech model, Muse Voice Transcribe, launches with 25 validated languages. Tajik, Uzbek and Russian are not among them. Neither are the languages of hundreds of millions of other people. And Meta Model API has no voice output at all.

So for a grandmother in Dushanbe, a builder in Samarkand, or a migrant worker reading a Russian contract at 2 a.m., "just ask AI" still means "first, learn English."

Sado closes that gap. It joins Meta's frontier reasoning with Meta's own open low-resource speech research, so anyone can talk to a real AI agent in their own language — and hear it answer back.

---

🧑‍🤝‍🧑 Meet Sado

Sado is not a chatbot with a microphone. It is a character — the same warm, witty, slightly old-fashioned companion wherever you meet it.

Trait	What it means
🗣️ Speaks like you do	Understands Tajik, Uzbek and Russian — and the way people mix them mid-sentence.
🧠 Remembers	Names, routines, promises, documents you showed it last week. You can see and delete every memory.
👀 Sees	Point your camera at a letter, a bill or a form. Sado reads it, explains it, and tells you what to do next.
🛠️ Acts	Saves facts and sets reminders through real tool calls, and asks before anything changes.
🎭 Has a personality	Consistent voice, humour and manners. Polite with elders, quick with students, precise at work.

---

🌍 One companion, four places

Sado runs on one brain and one memory, and switches modes to fit where you are.

🏠 At home — the family companion

Grandma: "Sado, Rustam lands on Friday at 6. Remind me on Thursday evening." Sado: "Of course. I'll remind you on Thursday at 7 in the evening."

Hands-free talk: speak, pause, and Sado answers and listens again. No buttons, no typing.

🙋 For you — the personal assistant

You: "Sado, remember my car insurance ends in March — and tell me two weeks before."

Notes, plans, reminders and personal knowledge — searchable by voice.

💼 At work — the agent

You (shares a contract PDF): "What do I have to pay, and by when?"

Brief, professional answers: Sado reads PDFs and Office files and pulls out the names, dates and amounts that matter.

🧭 On the go — the explainer

Migrant worker (photographs a Russian document): "What is this?" Sado (in Tajik): "This is a notice about your work permit. It expires on 15 November. Here is what to do…"

Camera-first: understand any paper, in any language, in seconds.

---

🌐 A living sphere

Sado is a 3D sphere drawn live on the GPU. While you talk its surface ripples with your voice; while it answers it turns violet and moves with its own; while it thinks its colours swirl. A hands-free conversation fills the phone's screen with the sphere, the last words said, and one button to stop.

💻 Sado on your computer

The same app runs on Windows as a quiet companion in the tray. Nothing sits on screen until you call it: say "Sado" or press Ctrl+Alt+S in any app and just talk. Only a slim bar with a live voice wave appears while Sado listens, thinks and answers, and it slips away when the conversation ends. The name is heard by a small Vosk model on the computer itself; nothing it hears leaves the machine.

Ask	Sado does
"What's on my screen?"	Takes a screenshot and explains it with Muse: an error, a chart, a form
"What is the agent in the terminal doing?" / "Tell it to run the tests"	Reads a terminal's latest output; types the next task into it
"Find my contract PDF" · "Open it"	Searches your own folders, newest first, and opens the file
"Pause the music" · "Louder"	Media and volume keys
"How's the computer doing?"	Battery, memory, disk space and load
"What did I copy?" · "Copy this"	Reads or fills the clipboard
"Minimize Chrome" · "Lock the computer"	Arranges windows, locks the screen
"Save it" (Ctrl+S) · "Run Get-Date" · "Close Notepad"	Shortcuts, PowerShell, closing windows — after your yes

• Looking is free, acting asks first. Anything that types, presses shortcuts, closes windows or runs commands waits for your yes, by voice or tap; each request is single-use, expires in five minutes, and the model can never confirm it itself.
• Only the owner may drive it. Computer control is off by default. The computer's own app carries a secret token, and a phone signed in with the owner's Google account counts as the owner's; anyone else reaching the server, even through a tunnel, can never control the computer.

📱 Sado runs your phone

On Android, Sado acts on the phone you are talking to, and does the thing rather than open a screen for it:

Ask	Sado does
"Call Mum"	Finds "Onam" or "Онам" in your contacts, asks, and places the call
"Text Aziz I'm ten minutes late"	Reads it back, asks, and sends the SMS (WhatsApp and Telegram open ready to send)
"Play Sherali Jo'rayev" · "Next song" · "Louder"	Starts music from a search; media keys and volume
"Wake me at 6:30" · "25-minute timer for the plov"	Straight into the phone's clock
"Meeting with the doctor tomorrow at 3" · "Email Aziz…"	Calendar event and email, filled in
"How much battery is left?"	Answers from the phone's own status
"Open Telegram" · "Way to the airport" · "Flashlight on" · "Wi-Fi settings"	Apps, maps, flashlight, settings pages, web search

• Calls and texts wait for your yes. Sado asks; it calls or sends only after "ha" / "бале" / "да" / "yes", by voice or tap.
• Your contacts stay on the phone. The server only hears the name you said; the app finds the number on the device, matching Latin and Cyrillic spellings.
• Honest when it can't. If the contact or app is not there, Sado says so instead of pretending.
• Open it talking. A quick-settings tile, an app shortcut, or holding the home button (with Sado as the phone's assistant) opens Sado already listening.

🔗 One Sado, every device

Sign in with Google and Sado is yours across devices. Everyone who signs in gets their own memories, reminders and lists, so a judge trying Sado never sees anyone else's. Signed in on both, the owner's phone can drive the computer — "Which windows are open on my computer?" — and the computer can reach the phone: say "set a timer on my phone" at the laptop and the phone's clock starts it.

If the laptop serving Sado is closed, the app moves to a backup server on its own and comes back once the laptop answers again. Without a GPU, Muse Spark Open on Baseten listens in place of Omnilingual ASR.

---

⚙️ How it works

flowchart LR
    U([🎤 Voice · 📷 Camera · 📄 File]) --> STT
    subgraph Hear
      STT[Speech-to-text<br/>Meta Omnilingual ASR<br/>language detected automatically]
    end
    STT --> B
    subgraph Think
      B[🧠 Sado agent<br/>Meta Muse Spark 1.3]
      B <--> M[(Memory<br/>SQLite full-text search)]
      B <--> T[🛠️ Tools<br/>memory · reminders · lists · web search<br/>phone · computer]
    end
    B --> TTS
    subgraph Speak
      TTS[Speech-ready text<br/>→ OmniVoice designed voices<br/>→ Meta MMS-TTS fallback]
    end
    TTS --> O([🔊 Sado's voice])

Loading

1. Hear — Meta's open-source Omnilingual ASR (LLM 1B) transcribes Tajik, Uzbek, Russian and English. Sado works out which language you spoke, so there is no language picker.
2. Think — Muse Spark answers as Sado, in a mode that fits the moment (Home, Personal, Work, On the go). It calls tools to save facts and set reminders, and reads photos, PDFs, Office files and recordings.
3. Remember — Facts are stored locally and recalled with full-text search tuned for Tajik and Uzbek word endings. The app shows every memory and lets you delete it.
4. Speak — Numbers, dates and money are written out the way each language says them, then spoken in one of six designed voices, in the mood you choose.

---

🧩 What runs where

Model	Role in Sado	By
Muse Spark 1.3 (Model API)	The brain: conversation, tool calls, reading photos and documents	Meta
Omnilingual ASR LLM 1B	Listening in all four languages	Meta FAIR, open source
MMS-TTS	Fallback voice when the main voice is unavailable	Meta, open source
OmniVoice	Sado's main voice: designed voices and moods	k2-fsa, open source (Apache-2.0)
Muse Spark Open (Baseten)	Listening when no GPU is around: hears Tajik, Uzbek, Russian and English	Meta
Vosk small model	Hears "Sado" on the computer, offline	Alpha Cephei, open source

Meta's hosted speech model, Muse Voice Transcribe, does not cover Tajik, Uzbek or Russian, and the Model API has no text-to-speech — which is exactly why Sado brings open speech models alongside Muse Spark.

---

🗺️ Languages

Language	Listen	Speak	Status
Tajik	✅	✅	Working
Uzbek (Latin)	✅	✅	Working
Russian	✅	✅	Working
English	✅	✅	Working
Any of 1,600+ Omnilingual languages	🔜	🔜	Pluggable language packs

Adding a language means adding a language pack — speech models, voice and normalisation rules — not rewriting Sado.

---

🔒 Trust by design

• You own the memory. View, edit or delete anything Sado remembers.
• Ask before acting. Calls, texts and anything that changes the computer wait for your yes; reminders and memories show up in the app.
• Yours alone. Sign in with Google and your memories, reminders and lists are kept apart from everyone else's.
• Knows its limits. No medical, legal or financial advice — Sado explains documents and points you to real help.

---

🚀 Run it yourself

Server (Python 3.12, uv, an NVIDIA GPU for the voices):

cd server
cp .env.example .env            # then set SADO_META_API_KEY from dev.meta.ai
uv sync --extra speech
uv run uvicorn sado.main:app --port 8000

Speech recognition worker (Linux or WSL, because fairseq2 has no Windows build):

bash server/scripts/download_models.sh
cd server/workers && ~/.venvs/sado-models/bin/python -m uvicorn omnilingual_asr_worker:app --port 8011

Computer control (optional, Windows): uv sync --extra desktop, then set SADO_DESKTOP_CONTROL=true and a random SADO_DESKTOP_TOKEN in server/.env, and build the desktop app with flutter build windows --dart-define=SADO_API_URL=http://localhost:8000 --dart-define=SADO_DESKTOP_TOKEN=<the same token>.

App (Flutter 3.44):

cd mobile
flutter run --dart-define=SADO_API_URL=http://<your-server>:8000

Checks: uv run pytest and uv run mypy src in server/, flutter test and dart analyze in mobile/.

---

🚧 Build log

Sado is being built live during the hackathon. All code in this repository was written from 27 October 2026, partly with Meta's own coding agent, Muse Code.

📐 Engineering principles · 🗺️ Roadmap · 🤖 Agent guide

• Foundations: server and mobile skeletons, Clean Architecture, CI
• Speak: six designed voices, six moods, speech-ready text in four languages
• Hear: Meta Omnilingual ASR for Tajik, Uzbek, Russian and English, language detected automatically
• Think: Muse Spark agent loop with tools, on the user's own clock
• Remember: memories saved by voice or by hand, recalled when relevant, visible and deletable
• See: explain photos, PDFs, Office files and recordings
• Act: reminders the agent sets and mentions when due
• Character: animated robot face that follows mood and activity
• Modes: Home · Personal · Work · On the go
• Live talk: hands-free conversation, silence ends the turn
• Computer companion: tray app, hotkey, voice wave; reads and drives windows and terminal agents, asking before it acts
• Hosted backend for judges
• Demo video: watch on YouTube

---

Sado — "voice" in Tajik and Uzbek.
Because everyone deserves an AI that speaks their language.