Your AI doesn’t just think — it speaks, listens, and remembers.
This is a Telegram-first voice layer that sends voice updates out, ingests voice notes in, and stores transcripts into your memory backend.

What it includes

• Telegram transport (outbound + inbound)
• Kokoro TTS via MUSE TTS (54 voices) for outbound voice notes
• faster-whisper STT for inbound voice transcription (local, fast, private-friendly)
• Pluggable transcript sink: local file, webhook, or MUSE Brain MCP adapter

What is Kokoro here?

Kokoro is the voice engine behind MUSE TTS. You get 54 voices and can pick a default persona voice without changing core logic.

What is faster-whisper here?

faster-whisper is a lightweight speech-to-text server you can run locally.
In this repo it exposes an OpenAI-compatible transcription endpoint, so the bridge can call it like any standard STT API.

Where transcripts go

Choose one with MEMORY_SINK_MODE:

• file (default): writes NDJSON to ./state/transcripts.ndjson
• webhook: POSTs transcripts to your own endpoint
• mcp: sends transcripts to MUSE Brain via mind_observe (optional adapter)

Vibe-coder friendly (entry level)

• You can run this with copy/paste setup + .env.
• No need to train models.
• Start local with sane defaults, then swap providers later.
• Works even if you keep your own stack and only want the Telegram+voice piece.

Quick start

cp .env.example .env
npm install
npm run build

By default transcripts are saved locally (MEMORY_SINK_MODE=file), so no MUSE dependency is required. Set TELEGRAM_CHAT_ID to lock ingestion to your chat (recommended default).

Start STT sidecar:

python3 -m venv .venv
source .venv/bin/activate
pip install -r stt/requirements.txt
python stt/faster_whisper_server.py

Start bridge:

npm run bridge

Send demo message:

npm run demo:notify

For the full setup flow see docs/QUICKSTART.md.