Fish Audio's Drama 3, a 35B text-to-speech model that whispers, laughs and shouts on cue, running in llama.cpp on one 24 GB card next to a small chat LLM, with a web page you talk to.
demo.mp4
demo/demo.mp4 is a short clip of it in use.
You need Docker Desktop with GPU support and an NVIDIA card with 24 GB
(RTX 3090, 4090 or 5090). From demo:
.\start.ps1
The first run downloads about 27 GB of models and builds the image, then prints "open http://localhost:8000". Press the round button and talk. Details, settings and the terminal version are in demo/README.md.
third_party/llama.cpp, a fork with the Drama 3 pieces: the converter to GGUF, the codebook tower as an mtmd graph with a persistent cache, the DAC codec, and a pipeline that takes reference clips and earlier turns as codes and streams audio as it decodes.demo/say.cpp, a small process that keeps the models loaded and renders a sentence per request, with a rolling history of what it last said so sentences join.demo/web, the page and its server: Parakeet TDT 0.6B v3 and Silero VAD on the CPU for hearing, Gemma 4 E4B for the answers, Drama 3 for the voice. The LLM puts a mood label in front of every sentence; the server turns it into a direction and picks a same-voice reference clip for it.DESIGN.md, the plan this follows, with what was measured along the way;bench/, the numbers;verify/, the checks against fish-speech.
The models themselves are not in the repo. The Drama 3 GGUFs are at EryriLabs/drama-3-llamacpp-GGUF; the download script fetches those, Gemma 4, Parakeet and Silero.
Mara speaks with the voice of speaker p283 from the CSTR VCTK Corpus
(University of Edinburgh), one sentence of which is the reference clip in
verify/refs/p283.codes; the mood clips in demo/bank were rendered from it.
VCTK is released under the Creative Commons Attribution 4.0 licence:
Yamagishi, Veaux and MacDonald, "CSTR VCTK Corpus: English Multi-speaker
Corpus for CSTR Voice Cloning Toolkit (version 0.92)", 2019,
https://doi.org/10.7488/ds/2645. If you ship the voice, keep the credit.
Fish Audio for Drama 3 and fish-speech, which was the ground truth for every check here; the llama.cpp project; NVIDIA for Parakeet and the sherpa-onnx team for its ONNX export; Unsloth for the Gemma 4 quants; the CSTR at Edinburgh for VCTK.
The code here is MIT. Drama 3 and the converted GGUFs are under the Fish Audio Community License, which allows research and non-commercial use freely and commercial use below a revenue threshold; read it before shipping anything.