How to Self-Host Ultravox with Docker: Step-by-Step Tutorial

Ultravox changes the voice AI game: instead of stitching together speech-to-text, an LLM, and text-to-speech, it is a single multimodal model that speaks directly. With roughly 4,500 GitHub stars and an MIT license, it is one of the most promising open-source voice models of the year. Here is how to run it yourself.

What You Need

• Minimum: 2 CPU cores, 4 GB RAM
• Recommended: 4 CPU cores, 8 GB RAM
• Software: Docker + Docker Compose

Step 1 — Pull the Image

The official image is published on GitHub Container Registry:

docker pull ghcr.io/fixie-ai/ultravox:latest

Step 2 — Run with Docker

Start the server and map port 8080 to your host:

docker run -d --name ultravox -p 8080:8080 \
  -v ./data/ultravox:/data \
  ghcr.io/fixie-ai/ultravox:latest

The volume keeps model data and configuration persistent across restarts.

Step 3 — Docker Compose (Production)

For production deployments, use a compose file so the service restarts automatically:

services:
  ultravox:
    image: ghcr.io/fixie-ai/ultravox:latest
    restart: unless-stopped
    ports:
      - "8080:8080"
    volumes:
      - ./data/ultravox:/data

Then start everything with docker compose up -d.

Step 4 — Connect Your App

Once running, your application can send text prompts to the Ultravox API and receive synthesized speech back in real time. This is what makes it so powerful for voice agents: you skip the transcription and speech-generation layers entirely and talk to a model that was trained to sound human.

Review Verdict

Ultravox is a standout for anyone building real-time voice experiences. The single-model design, open license, and clean Docker packaging make it the easiest path to production-grade voice AI on your own hardware.