Fish Audio's Drama 3 is an open text-to-speech model you can direct like an actor

Fish Audio released Drama 3, a text-to-speech model that takes plain-language acting direction. Instead of fiddling with sliders, you describe how a line should sound, which words to stress and how to deliver them, and the model performs it. It can also re-record a single word or phrase without regenerating the whole clip, which is a big deal for anyone editing dialogue.

Under the hood it is a 35 billion parameter mixture-of-experts model with about 3 billion parameters active per token. The weights are open on Hugging Face, and it is also available through Fish Audio's API. Self-hosting needs a single NVIDIA GPU with at least 48GB of memory in FP8; early users report that 4-bit versions sound close to the original.

Serving support arrived quickly. The vLLM-Omni project added Drama 3 for voice cloning and expressive speech, reporting 128 concurrent streams and a median of 255 milliseconds to the first audio on one H200. That makes it a realistic self-hosted option for audiobooks, dubbing and scripted dialogue.

Our take

Open-weight voice just caught up with direction, not just quality. The ability to fix one word is what moves AI voice from demos into real production workflows.