How much does it cost to self-host TTS?

Tens of euros a month per engine container for a small deployment, plus your time. The engineering, not the hosting, is the expensive half.

Which models would I run?

Piper for breadth and speed, Kokoro for naturalness, piper-plus for Japanese, an Indic model for Indian languages, faster-whisper for transcription - the same four Naadly runs.

Comparisons

Self-hosted TTS vs a hosted service, costed properly

Free models, paid afternoons: the comparison nobody does honestly.

Last checked 14 August 2026

In short

Running open TTS yourself costs roughly a small container per engine - tens of euros a month - plus the engineering that turns a model into a product: licence auditing, voice labelling, pronunciation handling, queueing sized to real memory, watermarking, consent records and being awake when a container OOMs. If you need one voice in one language, self-host. If you need a catalogue, several languages and an interface, hosting it is usually the more expensive option once your time has a number on it.

What the infrastructure really looks like

Naadly's own bill is the honest reference point, because it is exactly this architecture: separate containers per engine in Paris, sized from measured memory - a fast small one for Piper, a larger one for Kokoro because a single render needs about 700 MB, another for Japanese, a big one for Indian languages that is started on demand and stopped when idle, plus a database and object storage. That is a few hundred euros a month before anyone writes a line of application code.

The list of things that will bite you

A model swap while another request is mid-render can exceed the memory quota and kill every request on the container.
One shared voice cache will happily drop an entire language when an engine is briefly cold.
Watermarking libraries can carry thread-local state that breaks after the first request.
Half the open voices you find are non-commercial, and nothing tells you except the model card.
Concurrency has to be sized from measured RSS per render, not from CPU count.

When self-hosting wins

A single voice, a single language, an existing machine, no UI requirement, and data that may not leave your building. That is a real and common case, and Piper on a small container will serve it for years.

Questions

How much does it cost to self-host TTS?: Tens of euros a month per engine container for a small deployment, plus your time. The engineering, not the hosting, is the expensive half.
Which models would I run?: Piper for breadth and speed, Kokoro for naturalness, piper-plus for Japanese, an Indic model for Indian languages, faster-whisper for transcription - the same four Naadly runs.