Run it on your own hardware
Every model this service speaks with is open-weight and commercially licensed, and there is no third-party AI key anywhere in it. So the whole thing — catalogue, studio, API, queue, watermarking, consent records — can run inside your network, on machines you can point at. A hosted competitor cannot sell you that, whatever the contract says.
Who this is for
Three sorts of buyer, in our experience: the one whose scripts are not allowed to leave the building, the one whose per-character bill has stopped being funny, and the one who needs speech to keep working when the internet does not. Nothing about the product changes — a script written against the hosted API runs against your deployment by changing the base URL.
What a machine needs
No GPU. None of the engines need one, and none of the numbers on this page assume one — which is the part procurement usually does not believe.
- Reading only —
4 cores, 8 GB RAM, 40 GB disk
The catalogue, the studio, the API, the queue, projects, pronunciation, watermarking and provenance. - Reading and dubbing —
8 cores, 16 GB RAM, 100 GB disk
The above plus transcription, translation, subtitles and dubbing. - Everything —
16 cores, 32 GB RAM, 250 GB disk
The above plus Indian languages, cloning, performed delivery and the expressive engine.
The services, and the memory each one measured
Sizing a deployment by guess is how it fails in week two, so these are the figures from our own production measurements (status shows the same services running).
| Service | Engine | Memory | Needed |
|---|---|---|---|
| Web application and queue The shop, the studio, the API, the job queue and the provenance records. Object storage can be S3, R2, MinIO or a mounted disk. |
FastAPI, PostgreSQL, an object store | 1-2 GB | always |
| Reading voices 1,480+ voices at 21-43x realtime depending on how many slots the container is given; ~100 ms to first audio byte. This is the lane almost all volume goes through. |
Piper (MIT) | 512 MB-1 GB | always |
| Natural voices Heavier and more natural: ~700 MB per render, 8.4x aggregate realtime with two slots. |
Kokoro (Apache-2.0) | 1 GB | recommended |
| Japanese A separate service because the Japanese front-end - the G2P that makes Japanese worth shipping - lives in it. |
piper-plus (MIT) | 512 MB | if Japanese |
| Indian languages 22 Indian languages plus Indian English at 1.25x realtime, so Indic minutes are the expensive ones: about 0.8 CPU-hours per audio-hour. |
Indic-Mio Q8 (Apache-2.0) | 4 GB | if Indic |
| Expressive delivery Autoregressive, so about the length of the line on four threads, and 86 s to load. The qualified voices only. |
Chatterbox-Nano (MIT) | 8 GB | if expressive |
| Cloning and performed delivery torch floors this container at ~820 MB before a request arrives; peak 1.44 GB under concurrency. Consent records are enforced by the application, not by policy. |
kNN-VC, OpenVoice V2 | 4 GB | if cloning |
| Transcription and translation 3.6x realtime on English, 1.8x on Hindi, ~1 GB peak beside the two translation models. Hosted, this container is started per job and stopped again; on your hardware it can simply stay up. |
faster-whisper small, Marian/M2M | 2 GB | if dubbing |
What a licence includes
- Container images for every service above, or the source to build them.
- The voice catalogue with its per-voice licence record, so a legal team can audit what may be sold on and what may not.
- Watermarking, AudioSeal detection, consent records, synthetic-media disclosure and the provenance trail - the compliance surface is part of the application, not a hosted add-on.
- The HTTP API, the OpenAPI document, the Python and JavaScript SDKs and the MCP server, unchanged - a script written against the hosted product runs against your own deployment by changing the base URL.
- Deployment documentation (docs/self-hosted.md), the measured sizing above, and an upgrade path: images are versioned, and a new one adds the tables and columns it needs on boot rather than asking an operator to run migrations by hand.
- Named support by email, and a same-images-as-us guarantee: you are not running a fork that quietly stops getting fixes.
What it does not
- A GPU. Nothing in the stack needs one, and none of these numbers assume one.
- Third-party AI keys. There are none in the product, self-hosted or not.
- Hardware, and the electricity it takes. A deployment's cost is your own compute, which is the point.
- Voices whose training licence forbids redistribution - those stay hosted, and the licence record says which they are.
- A promise that expressive delivery covers your language. It is the 134 qualified voices today - English (UK) and English (US) - listed at /voices?expressive=1 - on your hardware exactly as on ours.
Why there is no price on this page
Because the honest number depends on things we do not know yet, and a headline price would have to assume the worst of them. It is quoted per deployment, on:
- Which of the services above you want - the expressive and Indic engines are the ones that cost real memory.
- How many people use it, and whether you resell what it renders.
- Whether the deployment is one site or many, air-gapped or connected.
- What support you need: email, or a response time with a number attached.
Hosted pricing is published in full on pricing, and most buyers should start there — a self-hosted deployment is worth it when the data boundary or the volume makes it worth it, not by default.
Ask for a quote
Kept to answer you and nothing else — see the privacy policy. Or write to [email protected] if a form is not your thing.