Fish Audio just dropped Drama-3 — a 35B MoE TTS model that’s worth wiring into audio.cpp.
It’s built for multi-character, long-form speech: up to 16 speaker tags in a single text, zero-shot voice cloning from a short reference, and point-tags like [sighs] or [laughs] for natural expression control.
Why it matters for us:
· State-of-the-art quality with a hybrid attention backbone and a codebook-MTP audio decoder that scales to long sequences.
· Production-grade serving: on a single H200 it hits sub-100 ms TTFA and stays real-time (RTF 0.80) even at 64 concurrent streams.
· Already SGLang-Omni ready, so there’s a clear deployment path we can mirror.
This is a strong candidate for the next-gen TTS backend — worth a look.
Yes! This model will be added. You can join the discussion here https://github.com/0xShug0/audio.cpp/discussions/891
PR merged! Please help test it and spread the word so more community members can collaborate on improving this model. Once the model is stable enough, I'll create a minor release to include it.
Up to 5× faster than real time at full speed, or 2× real-time speed with < 8 GB of VRAM on an RTX 5090.
Demo: