Deploy and Host a Whisper Speech-to-Text API on Railway

An OpenAI-compatible speech-to-text endpoint — POST /v1/audio/transcriptions — running Whisper on CPU through Speaches and faster-whisper. Point any OpenAI client at the URL, change the base URL, and keep the audio on your own infrastructure.

The transcription model ships inside the image, so the first request after the deploy returns a transcript instead of an error.

About Hosting a Whisper Speech-to-Text API

Speaches serves the OpenAI audio API on top of CTranslate2. Three things decide whether a deployment of it is usable, and they are the three this template fixes.

A model has to be present before the first call. Speaches does not fetch a model on demand — a deploy with an empty cache answers 404 Model 'Systran/faster-whisper-small' is not installed locally to every transcription request while the container stays healthy and /v1/models lists only the voice-activity detector. This template bakes Systran/faster-whisper-small into the image and copies it onto the volume on first boot, because a Railway volume is a bind mount that would otherwise hide anything the image build left under its path.

The thread pool should match the container. cpu_threads defaults to 0, which lets CTranslate2 size its pool from the cores it can see — the host's count, not the container's quota; this deploy's container saw 32 cores against a quota of 8. The entrypoint reads /sys/fs/cgroup/cpu.max and pins the pool per deploy. Measured on one host at --cpus 2 --memory 2g on the same 60.9-second file, warm: 15.03 s and 14.35 s with the default against 11.54 s and 11.69 s pinned. On Railway itself the same test is a smaller and noisier win — around 3.2 s pinned against 3.5 s unpinned — so this is a modest, not a dramatic, difference.

The endpoint has to be closed. Speaches installs its auth dependency only when API_KEY is set, so an unset key publishes transcription to anyone who finds the URL. The key is generated per deploy and the container refuses to boot without one. The Gradio playground, which is served with no authentication of its own, is off.

The volume is mounted as uid 0 by the platform while the image runs as ubuntu (uid 1000); the entrypoint repairs the ownership and drops privileges rather than running the server as root.

Common Use Cases

Transcribe recordings without per-minute billing — meetings, calls, voice notes and podcasts stay on your own instance.
Give an agent or bot ears — anything that already speaks the OpenAI audio API works by changing base_url.
Subtitles and search — response_format=srt, vtt, json or verbose_json from the same endpoint, in 99 languages, with /v1/audio/translations for English output.

Dependencies for Whisper Speech-to-Text Hosting

Speaches 0.9.0-rc.3 — OpenAI-compatible STT/TTS server (MIT)
faster-whisper / CTranslate2 — the inference engine
Systran/faster-whisper-small — baked into the image, int8 on CPU
A volume at /home/ubuntu/.cache/huggingface/hub for the model cache

Deployment Dependencies

speaches-ai/speaches
SYSTRAN/faster-whisper
bon5co/whisper-railway — the wrapper image used here, ghcr.io/bon5co/whisper-railway:0.9.0-rc.3

Implementation Details

API_KEY is generated per deploy as a Railway secret and is required on every /v1/* route. /health is the only unauthenticated route, because the platform's healthcheck dials it.
The deploy form asks for nothing. Every other setting has one correct value and lives in the image, since a template variable holding a literal publishes as a blank required field.
Larger models — Systran/faster-whisper-medium, …-large-v3 — install at runtime with POST /v1/models/Systran/faster-whisper-medium and persist on the volume. Kokoro TTS installs the same way if you want /v1/audio/speech from the same service.
$PORT is honoured, because Railway's HTTP healthcheck dials the injected port rather than the domain's target port.

What to expect on CPU: roughly five times faster than real time for faster-whisper-small on a two-vCPU container. Good for recordings and queues; a request long enough to exceed Railway's edge timeout should be split client-side. The volume belongs to one replica, so this service scales up rather than out.

Why Deploy a Whisper Speech-to-Text API 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 a Whisper speech-to-text API 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.