Fish Audio Drama 3: Directable 35B Open-Weight TTS
Fish Audio releases Drama 3, a 35B MoE text-to-speech model that follows bracketed performance direction, with open weights and a ComfyUI partner node.
[choked up, crying quietly] in front of a line changes how that line is read. ComfyUI's Fish Audio partner node gained a Drama 3 option on the same day in Comfy-Org/ComfyUI#16999.
What Drama 3 is
Drama 3 is a Mixture-of-Experts text-to-speech model with a 35B parameter text backbone and 3B active per token, using hybrid full plus linear attention. Audio is produced by a codebook-MTP decoder that decodes 10 codebooks causally over the whole sequence, and the expert weights are stored in FP8 with the remaining weights in bf16. The published package is eight safetensors shards plus a 1.9 GB codec, roughly 40 GB on disk.
Two things separate it from an ordinary voice-cloning model:
- Speaker tags.
<|speaker:0|>through<|speaker:15|>let a single text carry up to 16 characters, so a dialogue scene is generated as one stream instead of concatenated takes. The ComfyUI node marks speaker changes with@Voice1,@Voice2style tokens when several voices are connected. - Long-form handling. Long text is chunked automatically and can be streamed chunk by chunk, which is aimed at audiobooks, dubbing and audio drama rather than one-off sentences. Drama 3 speaks 83 languages.
Directing the performance
Direction goes in square brackets, written the way you would say it to an actor, and it can be placed at any point in the line so the read turns where the mood changes. Direction does not have to match the language of the line: the launch examples use Chinese, Spanish and Japanese direction on English text.
[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 <emphasis>...</emphasis> pair is a scoped tag: it directs only the words between the opening and closing tag, rather than the rest of the line.
 |  |  |
|---|---|---|
| Excited | Sad | Whispering |
Tone presets from the Drama 3 product page. Fish Audio demonstrates the same reference line read on S2.1 Pro and on Drama 3, and the differences land on the second direction in each line, which is where the voice breaks, goes quiet or gives way to a laugh.
Pronunciation robustness
Fish Audio reports Drama 3 as first on Artificial Analysis' Pronunciation Robustness benchmark, which tests hard words, numbers, codes and context-dependent readings inside English sentences and has human listeners check each one:
| What is being read | Drama 3 | Fish Audio S2.1 Pro |
|---|---|---|
| Names of brands, places and technical terms | 95.9% | 87.3% |
| Words whose reading depends on context | 95.3% | 86.1% |
| Shorthand: numbers, dates, units, notation | 93.6% | 92.7% |
| Exact sequences: codes, paths, emails | 86.3% | 79.4% |
| Overall | 93.2% | 88.1% |
The design change behind the numbers is in training: most speech models are tuned with rewards for being easy to transcribe and for sounding like the reference clip, which pushes a laugh or a whisper down. Drama 3 is tuned against a preference model over takes instead, so an expressive read is not penalised as a transcription error.
Running it
- Open weights. The full model is on Hugging Face, and Fish Audio documents running it through SGLang-Omni from launch day with
sgl-omni serve --model-path fishaudio/drama-3. - ComfyUI. The Fish Audio partner nodes in ComfyUI core ("Fish Audio Text to Speech") expose Drama 3 as a model choice as of October 13, 2026, alongside the existing
s2.1-proands1options. These nodes call the Fish Audio API, so they need an API key rather than a local GPU. - Hosted. The model is also live in Fish Audio's own Text to Speech app for web, iOS and Android.
Availability
Weights: fishaudio/drama-3 on Hugging Face. Deployment guide: Drama 3 Cookbook. Product page: fish.audio/drama-3.
Comments
Sign in with GitHub to join the discussion.