Drama 3: Open Weights Multi Speaker Dialogue TTS

Scratch dialogue shapes an edit before actors record anything. Most text to speech tools speak one voice at a time. You must render each line, then place it by hand. Drama 3 from Fish Audio (arXiv:2610.18010) reads a script with up to 16 speakers in one pass. It adds emotion and sound tags inside the text, clones voices from a short reference and lets you revise a line while it keeps the context of earlier takes.

Studio availability. Drama 3 is not available in AI FILMS Studio. This article covers an open weights release for local deployment. To generate speech in the cloud, use the voice workspace.

What the Model Card Says

Item	Detail
Size	35B total parameters, about 3B active per token (mixture of experts)
Attention	Hybrid of full and linear attention
Audio decoder	Codebook MTP decoder, 10 codebooks decoded causally
Precision	FP8 for the experts, bf16 for the rest
Languages	83 listed in the model tags
License	Fish Audio Community License

The card lists these features.

• Multi speaker scripts. Speaker tags <|speaker:0|> to <|speaker:15|> let one text hold up to 16 characters.
• Inline tags. Tags such as [sighs] and [laughs] place events at exact points in the line.
• Voice cloning. A request carries a reference audio clip and its transcript.
• Long text. The server splits long input into chunks and can stream each chunk.
• Editing by context. You can send earlier takes as previous_takes. A revised line then continues from them, which keeps pacing and tone. The card shows a sentence edited twice this way.

Speed

Fish Audio ran streaming tests on one NVIDIA H200 GPU with English prompts from the Seed-TTS test set.

Concurrent streams	Throughput (frames per second)	Median time to first audio	Real time factor (mean)
1	166.7	88.9 ms	0.131
16	927.6	210.1 ms	0.373
64	1,687.3	450.9 ms	0.801

A real time factor below 1 means the model makes audio faster than it plays. At 16 streams the card reports about 2.7 times faster than real time. Moving the vocoder to a second GPU at 256 streams raised throughput from 2,422 to 4,160 frames per second.

The card states no VRAM need. The 35B weights with FP8 experts are large, so a data center GPU such as the H200 is the tested setup. Treat this as a studio server model, not a laptop model.

Serve the Model

The model runs through two serving frameworks. Pick one.

# Option A: SGLang-Omni
git clone https://github.com/sgl-project/sglang-omni.git && cd sglang-omni
uv pip install --prerelease=allow -v -e ".[fishaudio]"
sgl-omni serve --model-path fishaudio/drama-3 --port 8000

# Option B: vLLM-Omni
git clone https://github.com/vllm-project/vllm-omni.git && cd vllm-omni
uv pip install -e .
vllm serve fishaudio/drama-3 --omni --port 8091

A request body has a text field, a references list and an optional previous_takes list. A script with two characters looks like this.

<|speaker:0|> You came back. [sighs] I thought the night shift was over.
<|speaker:1|> It is never over. [laughs] Pour me a coffee.

Read the framework documentation for the exact request path and the format of the reference audio. You can try the model in the public Hugging Face Spaces, including fishaudio/Drama-3.

Read the License Before Paid Work

The model page links the Fish Audio Community License. Its tag in the page metadata names a research license. These two names do not match, and the page does not state commercial limits. Open the license file in the repository and read it before you use any output in a paid project. Voice cloning also needs consent from the voice owner. Never clone a living actor without a written agreement.

Film Uses

• Scratch dialogue for a whole scene. Write the script with speaker tags and hear a table read in one pass.
• Animatic voices. Match tone with emotion tags and test timing against picture.
• ADR pre timing. Revise a line with previous_takes to see how a new wording changes the rhythm.
• Multilingual drafts. Test a dubbed line in another language before you hire a studio.

Related Voice Work

Other open voice models in our guides cover smaller setups. Read the open source voice cloning roundup, MiSoTTS, dots.tts and MOSS-TTS Local Transformer.

Frequently Asked Questions

How many speakers can one script hold?

Up to 16. Speaker tags run from <|speaker:0|> to <|speaker:15|>.

Which languages does it support?

The model tags on Hugging Face list 83 languages. Test each language you need, because the card gives no per language quality figures.

How fast is the first audio?

The card reports a median time to first audio of 88.9 ms for one stream on an H200 GPU. It rises to 450.9 ms at 64 concurrent streams.

What does voice cloning need?

A reference audio clip and its transcript in the request. Get written consent from the voice owner first.

How do I revise one line without losing the rhythm?

Send the earlier takes as previous_takes. The revised line then continues from them, which keeps pacing and tone.

Can I run it on a gaming GPU?

The card states no VRAM need and tests on H200 GPUs. The model has 35B parameters with FP8 experts, so a single consumer card is unlikely to work. Test before you plan a pipeline around it.

Voice in AI FILMS Studio

The voice workspace offers ElevenLabs Eleven V3 Text to Speech, MiniMax Speech 2.8 HD Text to Speech and MiniMax Voice Clone. Use them for finished dialogue in the cloud without a GPU server. In the Nodes Graph Editor you can send a voice track into a lip sync node.

---

Sources

Drama 3 Technical Report (arXiv:2610.18010) | fishaudio/drama-3 Hugging Face | Fish Speech GitHub (Fish Audio) | Fish Audio Community License