Skip to content

About

Fish Audio Drama 3 text to speech, natively in ComfyUI: voice cloning, dialogue and inline cues on cards from 8 GB, plus a native Qwen3-ASR. Built with Fish Audio.

Topics

Resources

Stars

187 stars

Watchers

1 watching

Forks

Repository files navigation

Write the script. It plays the part. Fish Audio Drama 3 in ComfyUI: a script with its directions in amber, playing.

ComfyUI-Drama3 runs Fish Audio Drama 3, a 35B text-to-speech model, natively in ComfyUI. Clone a voice from a few seconds of audio, cast up to 16 speakers in one script and write directions such as [sighs] or [whispers] where they should land. It runs on NVIDIA cards from 8 GB, with nothing to pip install.

Listen on the demo page · Weights on Hugging Face · Quick start · Reference

What's new

Fish Audio retrained Drama 3's weights on 2026-10-29. With them, 108 of 120 test takes come out complete, up from 69. The small files are now tuned on the model's own speech, decoding is faster on every card, long scripts render twice as fast with lines at once, and voice_guidance keeps two voices apart in a dialogue whose lines hear each other.

Before and after: complete takes 69 to 108 of 120, silent takes 48 to 1, FP8 on 24 GB 2.0x to 2.7x real time, W4A8 on 8 GB 0.6x to 1.0x, a ten-minute two-voice story 225 s to 111 s.

Important

Updating? Update the nodes first, then download a model file again. Older nodes refuse the new files with unexpected ['line_head.weight']. The first release's files moved to v1/ and still load in the new nodes. Saved voices keep working, and the codec file is unchanged.

Listen

Takes rendered with this release, cloned voices first, each shown as its script with the directions in amber. Click to listen on the demo page.

The cloned voices come from two public-domain films of 1948 and 1954 with uncredited narrators, and from open speech corpora recorded for research by anonymous volunteers (CSTR VCTK and Google FLEURS, CC BY 4.0). The demo page credits each one; the other voices are invented by the model.

Quick start

1. Install the nodes and restart ComfyUI. Search for Drama 3 text to speech in ComfyUI-Manager, or run comfy node install drama3, or clone the repository:

cd ComfyUI/custom_nodes
git clone https://github.com/PedroMarinhoDev/ComfyUI-Drama3

2. Download a model file, the codec and Qwen3-ASR, which writes the transcripts of voice clips. From the ComfyUI folder:

hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_fp8experts_bf16.safetensors --local-dir models/diffusion_models
hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_codec.safetensors --local-dir models/vae
hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_asr_1.7b_bf16.safetensors --local-dir models/asr

Swap the first file for a smaller one from Which file if your card or RAM is smaller.

3. Open a workflow and queue. Drag drama3_voice_clone.json into ComfyUI, pick a clip of clean speech in Load Reference Audio, write your text in Drama 3 Generate and run. drama3_text_to_speech.json needs no clip: the model makes up a voice.

To keep a voice, unmute Save Drama 3 Voice. It writes the encoded voice, a few KB, to output/drama3_voices, and Load Drama 3 Voice brings it back without the clip.

Which file

All four model files hold the whole model and complete the same share of our test takes. They differ in how the routed experts, 90% of the model, are stored. Speed is seconds of speech per second on an RTX 4090, so 1.0× is real time. The 12 GB and 8 GB columns ran with the rest of the card held by another process.

File Pick it for Size System RAM 24 GB 12 GB 8 GB
drama3_fp8experts_bf16 Fish Audio's own expert weights, on a 24 GB card 38.6 GB ~48 GB 2.7× 1.0× 0.6×
drama3_int8experts_bf16 The same quality on a smaller card 38.7 GB ~48 GB 2.7× 1.5× 0.9×
drama3_w6a8experts_bf16 6 GB smaller than FP8, almost as close to it as int8 32.7 GB ~40 GB 3.2× 1.3× 0.7×
drama3_w4a8experts_bf16 The fastest on any card, or 32 GB of RAM 24.6 GB ~32 GB 3.8× 2.0× 1.0×

You also need ComfyUI v0.40 or newer and about 1.6 GB for the codec. Quality and speed measurements are in the reference.

Writing a script

Directions are free natural language, so any wording works. Fish Audio's technical report (Appendix C) has the full syntax.

<|speaker:0|> [whispering] Don't move. I think it heard us.
<|speaker:1|> [laughs] It's only the cat. <whisper>Your face, though.</whisper>
<|speaker:0|> [sighs] Fine. You win.
Write To
<|speaker:N|> at the start of a line, N from 0 to 15 Give a line to a speaker
[sighs], [laughs], [pause] Put a sound or pause exactly there
<whisper>...</whisper> Shape only the words inside
[with bright eager excitement] at the start of a line Set the delivery of the whole line
Direction: icy calm. then Line: So this is goodbye. Give a director's note
[language: pt] after a line Hint the language; it is not spoken

Text in round parentheses is spoken. The model copies the reference clip's delivery: a flat clip gives a flat take.

Dialogue

The voice in Generate's slot voiceN speaks for <|speaker:N|>; a speaker without a voice gets one the model makes up. speakers on Drama 3 Generate sets how the lines are generated:

speakers How it generates Voices stay apart Lines react
one take per line (default) Each line on its own yes no
one take per line, hearing the line before Each line hears the take before it the second voice can drift, unless guided yes
one take per speaker All of a speaker's lines in one take, put back in script order yes no
together One take in which every line hears the others voices can blend, unless guided yes

In those two modes, with a voice for every speaker, voice_guidance 1.0 on Drama 3 Generate keeps the voices apart, at about twice the render time for those takes.

For a long script in one take per line, add Drama 3 Speed Options with lines_at_once 4: a ten-minute two-voice story renders in 111 s instead of 225 s. It is off by default because it changes what a seed generates.

Troubleshooting

"unexpected ['line_head.weight']" when loading a model file

The file is from the 2026-10-29 weights and the nodes are older. Update the nodes and restart ComfyUI.

A take falls silent after a second, or speaks only part of the text

Use a file of the 2026-10-29 weights and set top_k to 0: workflows saved before that fix keep the old 30. Then check the voice: the transcript must cover every word of the clip, and the clip should be clean speech cut between sentences. Load Reference Audio cuts at a pause and Qwen3-ASR writes the transcript, so the two match. A new seed usually helps too.

Speakers swap or blend voices in a dialogue

Give every speaker a voice and use one take per line, where voices cannot reach each other. To keep the lines reacting to each other, use hearing the line before or together with voice_guidance 1.0, at about twice the render time.

Out of memory, or the first run takes minutes

Lower expert_cache_gb on Load Drama 3, or pick a smaller file if system RAM is the limit. The first run reads the whole model file from disk once; later runs start in seconds. Pinning is not a way out: pinned experts cost the same system RAM as unpinned ones, only locked.

`voice_guidance` or `lines_at_once` does nothing

Both decode a frame as one CUDA graph, which needs the experts pinned where the kernels can read them. ComfyUI's --disable-pinned-memory does not decide that; pin_experts on Load Drama 3 does, and Generate names what is missing when it cannot batch. A cache holding under half the experts (int8 on a card under about 20 GB) leaves some of them to the CPU, which also stops the batching: raise expert_cache_gb or pick a smaller file.

The reference has every node and setting, how the model fits on an 8 GB card, the full measurements, how to build the files yourself and how close the port is to Fish Audio's code.

License and credits

The code is under the GNU GPL v3, except the parts ported from Fish Audio's code and Fish Audio's tokenizer and config, which remain under the Fish Audio Community License, and the parts adapted from Qwen's qwen-asr, under the Apache License 2.0. NOTICE lists them.

Note

Drama 3's weights and outputs fall under the Fish Audio Community License: free for research and non-commercial use, and for commercial use below USD 500k annual revenue; a hosted TTS service needs a license from Fish Audio. Don't clone the voice of an identifiable person without their consent.

Built with Fish Audio. Drama 3 and its codec are by Fish Audio; Qwen3-ASR is by the Qwen team. This project is not affiliated with or endorsed by either.

About

Fish Audio Drama 3 text to speech, natively in ComfyUI: voice cloning, dialogue and inline cues on cards from 8 GB, plus a native Qwen3-ASR. Built with Fish Audio.

Topics

Resources

Stars

187 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages