Deploy and Host VoiceStudio on Railway

VoiceStudio is an open-source alternative to ElevenLabs. It clones a voice from a short sample, designs new voices from a description, generates speech in hundreds of languages, dubs videos, transcribes audio and builds audiobooks. This template runs the official VoiceStudio server image, so you get the full Studio web UI in your browser plus an OpenAI-compatible speech API. The models run inside the container, so your text and voice samples never go to a third-party API.

About Hosting VoiceStudio

It's one service and a Railway volume.

Locked from the first boot. A random admin key (OMNIVOICE_API_KEY) is generated at deploy time. The web UI asks for it once and swaps it for a short-lived browser session; API clients send it as a Bearer token. Without it, everything except the page shell is refused.
HTTPS out of the box. Railway terminates TLS, so the key and session cookie never travel in plain text.
Models and voices survive redeploys. The project database, your cloned voices, settings and the Hugging Face model cache all live on the volume at /app/omnivoice_data.
Lean on memory. Cached files the server isn't using are dropped every 5 minutes, because Railway bills the cache as RAM: about 600 MB idle before the first generation instead of 2 GB.
CPU inference, tuned. Railway has no GPUs, so VoiceStudio runs on CPU. The template caps PyTorch at your plan's vCPUs (OMP_NUM_THREADS, default 8). Without that cap PyTorch sees the host's 48 cores, oversubscribes the 8 it's allowed, and a single sentence took over 15 minutes in our tests; with it, a short sentence takes about 20 seconds.

Common Use Cases

Voice-overs and narration for videos, courses and podcasts without per-character pricing
Cloning your own voice (or a consenting speaker's) for consistent narration across projects
A private, OpenAI-compatible text-to-speech endpoint for your own apps and agents

Dependencies for VoiceStudio Hosting

Nothing else: models download from Hugging Face on first use and are cached on the volume
Optional: a Hugging Face token (HF_TOKEN) for gated models

Deployment Dependencies

Implementation Details

Sign in. Open the service's Railway domain. When the UI asks for a key, paste OMNIVOICE_API_KEY from the Variables tab. The browser keeps only a short-lived session, not the key.

First run. The first generation downloads the default model (about 3.4 GB) into the volume and loads it, so it takes a minute or two. After that, a 4-second clip takes about 20 seconds on the Hobby plan's 8 vCPUs. Long texts, audiobooks and video dubbing take proportionally longer, so run them from the web UI rather than through a single HTTP request (Railway closes HTTP requests after 5 minutes).

Use the API. Send the key as a Bearer token:

curl https://.up.railway.app/v1/audio/speech \
  -H "Authorization: Bearer " \
  -H "Content-Type: application/json" \
  -d '{"model": "tts-1", "input": "Hello from Railway", "voice": "alloy", "response_format": "wav"}' -o hello.wav

Apps in the same Railway project can call http://${{VoiceStudio.RAILWAY_PRIVATE_DOMAIN}}:3900 instead, with the same key. That traffic isn't billed as egress.

Resources and cost. The image is about 6 GB (it ships PyTorch), so it needs the Hobby plan or higher. Railway bills cached files as memory, so every 5 minutes the template drops cached files the server isn't using; that saves about 1.5 GB. Measured on Railway: about 600 MB of RAM before the first generation and about 4.8 GB once a model is loaded. VoiceStudio keeps most of that memory even after it unloads an idle model, so at Railway's memory pricing a used instance costs around $48 a month if left running. Remove the service (or scale it down) when you're not using it. The default model takes about 3.4 GB of the 5 GB Hobby volume; grow the volume (Pro) or remove models you don't use from the Model Catalogue before installing more.

Updates. The template pins a VoiceStudio release. To upgrade, change the image tag on the service to a newer ghcr.io/debpalash/voicestudio: and redeploy; your data stays on the volume.

Telemetry. Product analytics are opt-in and off unless you accept the first-run prompt. Set OMNIVOICE_ANALYTICS_DISABLED=1 to remove the prompt entirely.

Use voice cloning responsibly. Only clone voices you own or have explicit permission to use.

Why Deploy VoiceStudio on Railway?

Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.

By deploying VoiceStudio on Railway, you are one step closer to supporting a complete full-stack application with minimal burden. Host your servers, databases, AI agents, and more on Railway.