Is Naadly open source?

The models behind the voices are open and named, with their licences printed. The service around them is not.

Which open TTS model is best?

Piper for speed and breadth, Kokoro for naturalness, piper-plus for Japanese, the Indic family for Indian languages. There is no single winner, which is why Naadly runs four.

Can I run these models myself instead of paying?

Yes, and for a single-voice single-language need you probably should. The maths changes as soon as you need licences, labels, editing and uptime.

Guides

Open-source text to speech: what it costs to run yourself

The models are free. The afternoons are not.

Last checked 2026-08-14

In short

Open text-to-speech models - Piper, Kokoro, piper-plus and the Indic family - are genuinely production-grade for narration, and running them on a small CPU container is cheap. The cost is everything around them: licence checking per voice, pronunciation handling, memory limits, queueing, watermarking, and the week you spend discovering a model swap can OOM a container. Naadly is those models with that work already done.

What the open models are actually like

Piper. Fast, tiny, unfussy - many times realtime on a CPU. Neutral rather than expressive; excellent for narration and phone systems.
Kokoro. Noticeably more natural, heavier, needs about 700 MB per render - which is why a 1 GB container fits exactly one at a time.
piper-plus. Where the Japanese front-end lives, and the reason Japanese is worth shipping at all.
Indic models. Twenty-plus Indian languages, and about 1.25x realtime: correct, and expensive in machine time.
faster-whisper. Transcription, roughly 10x realtime on English.

The work that is not the model

Everything we had to build after the models worked: a licence audit per voice so the catalogue can be sold commercially; labels for gender, accent, pace and quality, measured rather than guessed; a pronunciation layer; per-line re-rendering; a queue sized to real container memory; separate caches per engine so a cold container cannot delete a language from the shop; watermarking that survives concurrency; consent records for cloning. Each of those was a real bug before it was a feature.

When self-hosting is right

One language, one voice, no editing UI, a machine you already pay for, and an appetite for maintenance - self-host, and do not let anybody sell you otherwise. When you need many voices with clean licences, an interface a non-engineer can use, several languages, cloning with consent records, or somebody else on call at 2am, hosting it yourself stops being the cheap option.

Why this matters for the price

Because Naadly runs these models on its own containers in Paris, the marginal cost of a render is machine time rather than a per-character fee to somebody upstream. That is what makes 30 free minutes and uncapped standard speech on $39 possible at all - and it is also why Indic minutes are metered, because that model really is expensive to run.

Questions

Is Naadly open source?: The models behind the voices are open and named, with their licences printed. The service around them is not.
Which open TTS model is best?: Piper for speed and breadth, Kokoro for naturalness, piper-plus for Japanese, the Indic family for Indian languages. There is no single winner, which is why Naadly runs four.
Can I run these models myself instead of paying?: Yes, and for a single-voice single-language need you probably should. The maths changes as soon as you need licences, labels, editing and uptime.