Deploy and Host Kokoro TTS Lite on Railway

Kokoro TTS Lite is an OpenAI-compatible text-to-speech API for the Kokoro-82M model, running on ONNX Runtime instead of PyTorch. It answers POST /v1/audio/speech exactly like OpenAI's endpoint, so the official OpenAI SDKs work against it with nothing changed but the base URL — and it fits inside Railway's smallest memory limits, including the 0.5 GB Free plan.

About Hosting Kokoro TTS Lite

The usual Kokoro server images run PyTorch and hold about 1.1 GB resident, which is more memory than the Free (0.5 GB) and Trial (1 GB) plans allow — the container is killed during model load and never serves a request. This template runs the ONNX build and ships both quantised and full-precision weights, so one variable picks the tradeoff that matches your plan:

`MODEL_PRECISION`	Resident memory	11 seconds of audio	Runs on
`int8` (default)	~375 MB	~9 s	Free, Trial, Hobby
`fp32`	~610 MB	~3 s	Trial, Hobby

Both numbers are measured, not estimated. The model weights are baked into the image rather than downloaded at boot, so a cold start is a few seconds and does not depend on your network. A healthcheck on /health means Railway only routes traffic once the model is actually loaded.

Why Deploy Kokoro TTS Lite 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 Kokoro TTS Lite 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.

Common Use Cases

Drop-in replacement for OpenAI's text-to-speech API, at no per-character cost
Voice for AI agents and chatbots that already speak the OpenAI protocol
Narration for video, podcast, and course pipelines
Accessibility read-aloud for web and mobile apps
Notification and IVR audio generated on demand

Implementation Details

from openai import OpenAI

client = OpenAI(base_url="https://your-app.up.railway.app/v1", api_key="unused")

client.audio.speech.create(
    model="kokoro",
    voice="af_bella",
    input="Hello world!",
).stream_to_file("output.mp3")

API Endpoints:

Endpoint	Purpose
`POST /v1/audio/speech`	OpenAI-compatible speech synthesis
`GET /v1/audio/voices`	List the available voices
`GET /v1/models`	Model listing, for clients that probe it
`GET /health`	Healthcheck

Variables:

Variable	Default	Meaning
`MODEL_PRECISION`	`int8`	`int8` for the Free plan, `fp32` for ~4x the speed
`DEFAULT_VOICE`	`af_heart`	Voice used when a request does not name one
`API_KEY`	unset	When set, `/v1/*` requires `Authorization: Bearer`
`OMP_NUM_THREADS`	`2`	ONNX thread cap; raise it to match your plan's vCPU count
`PORT`	`8880`	Listen port

Audio formats: mp3, wav, pcm, opus, aac, flac. 50-odd voices across English, French, Italian, Japanese, and Chinese, and voices can be blended by name.

Need the full PyTorch feature set — streaming with configurable chunking, word-level caption timestamps, weighted voice combination endpoints, the bundled web player? Deploy Kokoro TTS API instead. It needs 1.5 GB of RAM and so requires the Hobby plan or above.

Dependencies for Kokoro TTS Lite Hosting

Docker container runtime
Kokoro-82M model weights, Apache 2.0 licensed

Deployment Dependencies

Source repository for this template
kokoro-onnx — ONNX runtime bindings and weights
Kokoro-82M on Hugging Face