GitHub radar
Text-to-speech for multi-character scenes and long-form content
Fish Audio Drama 3 is a next-generation text-to-speech model built for multi-character dialogue and long-form text. It voices up to 16 speakers in one pass, no manual track-splicing needed.
It's a TTS model (turns text into audio) that switches voices using speaker tags in the text. It has 35 billion parameters total, but only 3 billion are active per token (a small chunk of text).
When it helps
Useful if you're making videos, podcasts, or audio drama with several speaking characters and want a ready scene without splicing different voices together. Skip it if one voice or a ready-made cloud TTS service is enough.
Pros
- Up to 16 voices in one text via speaker tags
- Clones a voice from a short audio sample
- Renders audio faster than real time even at 16 requests
- Free for non-commercial use and revenue under $500,000
Cons
- No ready Ollama/llama.cpp build, needs a separate server
- Needs a 48GB GPU or a Mac with 64GB memory
- Setup via SGLang-Omni or vLLM-Omni is more involved than typical TTS
How to set it up — step by step
- 1Open the model page on Hugging Face: https://huggingface.co/fishaudio/drama-3
- 2Check that you have a 48GB+ GPU or a Mac with 64GB+ memory
- 3Open your AI agent (Claude Code, Codex) in your project folder
- 4Send the agent the text from the block below
- 5Wait for the agent to deploy SGLang-Omni or vLLM-Omni and build a test scene
- 6Listen to the scene with two characters and check the voice switching by tags
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the Fish Audio Drama 3 text-to-speech model (https://huggingface.co/fishaudio/drama-3) on my machine through SGLang-Omni or vLLM-Omni for my GPU, and generate a test two-character dialogue scene.
The model weighs about 41 GB on disk, so you need a GPU with 48 GB of memory or a Mac with 64 GB or more. There is no Ollama build — you set it up through the SGLang-Omni or vLLM-Omni server.
Open on Hugging Face▌ More finds