Drama 3: an open-weight voice you can actually direct

Most voice models that can act are closed. You send text to an API, you get audio back, and that is the whole relationship. Fish Audio’s Drama 3 breaks that pattern. The weights are on Hugging Face, and the model takes direction in plain language: “say this like you just heard bad news”, “whisper”, “laugh, then answer”. We benchmarked it the day after launch, and it is now available on Speko.

Why open weights matter for voice

A closed voice model is a rented voice. The provider can change it, reprice it, or retire it, and your product changes with it. Open weights give a team three things an API cannot.

You can run it where your data lives. Drama 3 ships in BF16 and FP8 and runs on SGLang and vLLM. A team with strict data rules can keep every word of the script on its own hardware.

You can fine-tune it. A game studio can teach it a character. An audiobook house can teach it a narrator. That work stays with the team that did it.

You can see what you are building on. It is a 35B mixture-of-experts model with about 3B parameters active per token. That is why it is fast: only a small slice of the model runs on each step.

Open weights with this level of acting control are rare. Most open TTS models read text well and stop there. Drama 3 performs.

What we measured

We ran Drama 3 against Fish’s previous flagship, s2.1-pro, with the same voices, texts and regions.

• Naturalness. Drama 3 scored higher than s2.1-pro on all five voices we tested.
• Speed. On Fish’s hosted API from US East, first audio arrived in about 172ms, roughly 70ms faster than s2.1-pro in the same runs.
• Direction. We gave it 24 free-text emotions, from “furious” to “nostalgic”. In blind tests it audibly changed delivery for 92% of them, and 23 of 24 were usable. Sighs and laughs landed every time.
• Price. $10 per million bytes of text, down from $15 for s2.1-pro.

Direction did not cost quality. Directed clips scored as natural as plain reads.

Where it fits

Drama 3 is built for content you render once and play many times: audiobooks, game dialogue, trailers, explainers. Those are also the teams most likely to want their own copy of the model, so the open weights fit the use case.

For live phone agents, keep the usual check in mind and test it on your own script first, especially lines with numbers and symbols.

License, in one paragraph

Drama 3 uses the Fish Audio Community License. Commercial use is free while your company makes under $500,000 a year, and above that you contact Fish for a license. Offering it as a hosted service needs a separate agreement, and outputs cannot be used to train other foundation models. Self-hosted products must show “Built with Fish Audio”. If you only want to call it, the hosted API and Speko need none of this.

Try it

Drama 3 is live on Speko as fishaudio:drama-3. Pick it as your TTS model, or pin it per call next to your other providers. If you want the weights, they are at huggingface.co/fishaudio/drama-3.

Congrats to the Fish Audio team. More of this, please.