Unmute
An open, low-latency cascaded voice stack: Kyutai streaming STT plus any OpenAI-compatible text LLM plus Kyutai streaming TTS. The practical way to give a smart self-hosted LLM a natural real-time voice.
Overview
Best for: Self-hosting a voice front-end for your own (possibly private) LLM with good latency.
At a glance
Cascaded stack. 750 ms is TTS latency on one L40S, about 450 ms with services on separate GPUs. Code MIT, STT/TTS weights CC-BY-4.0; LLM licence depends on your choice. English and French speech.
Browser mic via the bundled frontend
Streamed TTS audio
English and French for STT/TTS (per model names); LLM language depends on your model.
Kyutai TTS voice set (kyutai/tts-voices).
Vendor: TTS latency about 750 ms on a single L40S, about 450 ms with services on separate GPUs.
Self-hosted.
Self-hosted; your responsibility.
CUDA GPU with at least 16 GB VRAM for the default setup; three or more GPUs recommended to split STT, TTS and LLM for lower latency.
Unmute code MIT; Kyutai STT/TTS weights CC-BY-4.0; LLM licence depends on your choice.
Features
- plug in any OpenAI-compatible LLM
- semantic voice activity / turn-taking
- voice selection
- Docker Compose and Swarm deployment
Pricing
No public price list.
No licence fee; you pay for GPU time. Plus LLM cost if you use a paid LLM endpoint.
n/a (self-hosted)
Free tier: Open weights
Source: github.com
Setup
- Linux x86_64 with a CUDA GPU (16 GB is enough for the default Gemma 3 1B setup).
- Set your Hugging Face token.
- Run docker compose; optionally point KYUTAI_LLM_URL at your own OpenAI-compatible LLM.
Endpoint
Local web app served by the compose stackAuthentication
None by default.
Quick start python
git clone https://github.com/kyutai-labs/unmute && cd unmute
export HUGGING_FACE_HUB_TOKEN=hf_...
# optional: use your own LLM (Ollama, vLLM, OpenRouter...)
# export KYUTAI_LLM_URL=http://host.docker.internal:11434
# export KYUTAI_LLM_MODEL=llama3.1
# export KYUTAI_LLM_API_KEY=...
docker compose up --buildWritten from the current docs. Check the vendor's SDK version before you ship.
Warnings
Cascaded, not native
Tone and emotion in the user's voice are not passed to the LLM; it trades that for a much smarter brain than Moshi.
No tool calling out of the box
Kyutai suggests wrapping vLLM in your own server to return tool-call results; Unmute itself is tool-unaware.
Platform restrictions
x86_64 Linux only (or WSL); no Mac or ARM servers.
English/French speech only
STT/TTS models cover English and French.
GPU count drives latency
Single-GPU deployments roughly double TTS latency versus split GPUs.
Plus 3 warnings that apply to all open models APIs. See category warnings.
Limits
- No built-in tool calling (wrap your LLM server to add it)
- Linux x86_64 only (Windows via WSL); no aarch64 or native Mac
Models and products
| Name | Status |
|---|---|
| kyutai/stt-1b-en_fr, kyutai/stt-2.6b-en | GA |
| kyutai/tts-1.6b-en_fr | GA |
| Any OpenAI-compatible LLM (default Gemma 3 1B; vLLM, Ollama, OpenRouter) | GA |
Docs and sources
Docs
Sources used
Exact environment variable values for Ollama in the snippet; full language list.