Open source narration infrastructure

Let every page speak.

Auritus turns the readable text on any page into on-demand audio — with a small embed, open-model workers you run, and an AWS safety net behind them.

Start with the embed See the architecture

Embed: script tag on a publisher page
Local + cloud: compute paths: local first, cloud fallback
Open models: models and deployment choices

Hear it for yourself

What Auritus does

A 30-second introduction, narrated by Auritus

Press play. What you hear is this page reading itself.

Auritus turns any article into audio on demand, from a single script tag — no pre-recording, no studio, no per-article cost.

You choose the voice model. You decide where the text goes. Your own GPU does the work, and AWS picks up the slack when it cannot.

It's open source, and every article on your site can speak, with the choices that matter still yours.

Narrated by Auritus with Kokoro (af_heart) — the same embed documented in Usage.

What just happened

One script tag, and the page reads itself.

Your page carries one script tag. Auritus reads the page text, audio is rendered on demand and returned as private audio behind a short-lived link, and a reader listens to it on the same page.

Control is the product

A narration layer that adapts to your stack.

Build a listening experience without treating your publisher content and production choices as somebody else's defaults.

Your data path

Narration on infrastructure you choose.

Keep page text in a stack you operate. Audio stays private by default and is delivered through short-lived access.

Your model

Use an open model. Tune the tradeoff.

Choose the voice backend and parameters that fit your quality, latency, and cost targets instead of accepting one fixed service.

Your compute

Start local. Fall back when it matters.

A GPU worker you run can claim ordinary work first. AWS Batch is there when that worker is unavailable or the queue needs help.

Designed for a reliable handoff

Your GPU when it is there. AWS when it is not.

A local worker races to claim the job. If nobody claims it during the configured window, a workflow submits an AWS Batch GPU job. The conditional claim means only one worker finishes the audio.

Content-hash deduplication keeps repeated requests efficient.
Private audio is returned with short-lived access.
Per-site quotas, budgets, and a kill switch limit cloud work.

Explore deployment options

The page asks for audio. A single conditional claim means only one worker ever wins it. Your own GPU claims it first; AWS Batch renders it only if no claim arrives before the timeout. Either path produces private audio, returned to the page as a short-lived link.

A short path to first audio

Deploy, register, embed.

Deploy your stack
Provision the API, worker fallback, private audio storage, and operator authentication in AWS.
Register an origin
Create a site key scoped to the publisher domain that will host the player.
Drop in the player
Add the embed once. Auritus extracts the readable content and keeps the player in step with your page.

Read the self-hosting guide

Security architecture

No IAM keys on the worker. None to leak.

The usual way to let a machine call your service is a long-lived IAM access key, scoped wider than anyone intended and rotated on a schedule someone has to remember. Auritus issues none. An operator signs in once — with a Cognito account, or your own identity provider — and the worker then carries a one-hour token that renews itself and cannot call AWS APIs at all. The risk is removed by the architecture rather than managed by procedure.

The security page states plainly which parts are enforced today and which are still being built, including the remaining JWT validation work.

Read the authentication and security design