Introduction

Last month we covered Drama 3 as an early-access preview. On October 13 (just after midnight MYT on the 14th), Fish Audio shipped it in full. Drama 3 is a text-to-speech model that takes stage directions in plain language. It's live today on fish.audio, in the Fish Audio iOS and Android apps, and in the API under the drama-3 model header. The bigger news is that the weights are open: fishaudio/drama-3 is up on Hugging Face, and SGLang-Omni can serve it from day one. Fish announced a $52M seed round alongside the launch.

The launch post puts it plainly: "Drama 3 is open-weight rather than open source." Read the license before you build a business on it.

What Changed Since the Preview

  • Direction in brackets, anywhere in the line. For example: [barely holding it together, forcing a smile] I'm fine. [shaky breath] Really. When you add a new direction mid-line, the read changes at that point. Paired word tags such as <whisper>…</whisper> limit a direction to just the words inside them.
  • 83 languages, and the direction doesn't have to be in the same language as the line. You can write a Chinese stage note over an English sentence.
  • Up to 16 speakers in one script via <|speaker:0|>…<|speaker:15|> tags. Long text is chunked and streamed automatically.
  • Pronunciation. Fish cites 93.2% on Artificial Analysis' Pronunciation Robustness benchmark (#1 on that board), up from 73.2% for S2.1 Pro. The biggest gains are on numbers, units, file paths and emails.
  • Coming soon: interactive "softer than that take" re-reads and one-word audio inpainting are still invite-only. An editing API is coming but isn't public yet.

Under the Hood

  • 35B-parameter MoE, about 3B active per token (8 of 256 routed experts plus one shared). The backbone was initialized from Qwen3.5-35B-A3B's language model.
  • Hybrid attention: 40 layers, in blocks of three Gated DeltaNet layers to one full-attention layer. Only one layer in four keeps a KV cache, which helps with long scripts.
  • Codebook-MTP audio decoder replaces S2's per-frame Fast AR. It uses S2's 44.1 kHz, 10-codebook DAC-based tokenizer.
  • Tuned with a single learned human-preference reward (GRPO, no KL penalty) instead of transcription and likeness scores. Fish says that's why a laugh you ask for no longer gets scored down.
  • Weights: FP8 MoE experts with the rest in bf16, about 40 GB on disk across 8 safetensors shards. Fish's serving numbers are on one H200: ~89 ms time to first audio and RTF 0.13 for a single stream, with RTF still under 1.0 at 64 concurrent streams.

Where to Run It Today

Self-host (SGLang-Omni):

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

The Drama 3 cookbook covers voice cloning from a reference clip, streaming, and putting the vocoder on a second GPU. Fish's code is tagged fish-speech v3.0.0-beta, which is marked as a pre-release.

API: it's the same POST https://api.fish.audio/v1/tts endpoint as S2.1 Pro. Add the header model: drama-3. Launch pricing is $10 per million characters through November 12 (80% off the $50 list price).

ComfyUI: the Fish Audio Text to Speech partner node (partner/audio/Fish Audio) now has a drama-3 model option, added in ComfyUI PR #16999. It supports bracketed directions, word tags and @Voice1/@Voice2 multi-speaker dialogue. Like the other partner nodes, it runs through the Fish API, not your local GPU.

Fish lists other day-one partners too, including OpenRouter, Runway, Pika, OpenArt, Vercel and Runware.

Paste-ready line to try:

[breathless, speaking fast] Just turn off the lights. Stay away from the windows. [voice drops to a terrified whisper] And whatever you hear out there… [slow, deadly serious] <emphasis>do not open that door.</emphasis>

The License Fine Print

Drama 3 uses the new Fish Audio Community License v1.0:

  • Free for research and non-commercial use.
  • Free for commercial use while you and your affiliates make under US$500,000 a year.
  • Running Drama 3 as a hosted TTS, voice-cloning or inference API for others always needs a separate license from Fish.
  • Making finished voiceovers or audiobooks for clients is allowed.
  • Cloning a real person's voice without consent is banned. So is removing watermarks, and so is using outputs to train other speech models.

S2 stays under the older Fish Audio Research License.

Conclusion

The preview showed that directable TTS was coming. Now you can download it. Indie audio-drama makers, game teams and audiobook narrators under the revenue cap can run a 35B emotive TTS model on their own hardware. Everyone else can try it in ComfyUI or through the discounted API before November 12. Start with one scripted line with two mid-sentence direction changes and listen for where the read turns. That's what this model is built to get right.

—Anabel