This week, Fish Audio released Fish Audio Drama 3. Drama 3 is a text-to-speech model from Fish Audio that takes plain-language acting direction. It is a 35B mixture-of-experts model with about 3B parameters active per token. The weights are on Hugging Face, and the model is also live in the Fish Audio API. For developers, it is a self-hostable option for scripted dialogue, audiobooks and dubbing.
TL;DR
- Size: 35B total parameters, about 3B active per token (8 of 256 experts plus 1 shared). Context length: Not disclosed.
- Runs on: 1 NVIDIA GPU with at least 48 GB in FP8. No CPU, Apple MPS or AMD ROCm. Official benchmarks use 1 H200.
- Performance: Fish Audio’s highest-scoring model on Artificial Analysis, but still behind several closed APIs on quality Elo.
- Best: 95.9% on brands, places and technical terms in Artificial Analysis Pronunciation Robustness.
- Worst: 86.3% on exact sequences (codes, paths, emails), its weakest pronunciation category.
- Bottom line:
- Best: directable, multi-speaker speech with open weights.
- Worst: 48 GB GPU floor and a revenue-capped license.
What is Fish Audio Drama 3?
Drama 3 turns a script along with bracketed directions into acted speech. Writers place cues like [barely holding it together] before a line. Paired tags such as
The model supports 83 languages. Speaker tags <|speaker:0|> to <|speaker:15|> allow up to 16 characters in 1 request. The API takes 1 voice ID per speaker, so a full scene renders in a single call.
How does the Drama 3 architecture work?
According to the launch post, the backbone starts from the Qwen3.5-35B-A3B language model. It has 40 layers, grouped as 3 Gated DeltaNet layers per 1 gated full-attention layer. Only 1 layer in 4 keeps a KV cache. DeltaNet layers store history in a constant-size state instead.
A new Codebook-MTP decoder replaces the per-frame Fast AR decoder used in S2. It has 4 shared layers feeding 9 layers, 1 per acoustic codebook. Only codebooks 0 and 1 feed back into the backbone; S2 fed back all 10.
The audio tokenizer is unchanged from S2. It is built on the Descript Audio Codec and outputs 44.1 kHz audio. It runs at about 21.5 frames per second.
Training ran 3 pre-training stages, then instruction tuning, then reinforcement learning. RL used GRPO without a KL penalty and a single learned human-preference reward model.
‘);});
var elo=[[‘Eleven v4 Turbo’,1328,’b’],[‘Qwen-Audio-3.0-TTS-Plus’,1266,’b’],[‘Fish Audio Drama 3 Preview’,1242,’a’],[‘Fish Audio S2.1 Pro’,1190,’b’],[‘Fish Audio S2 Pro’,1116,’b’]];
var ew=R.querySelector(‘#d3elo’);elo.forEach(function(r){ew.insertAdjacentHTML(‘beforeend’,”);});
function bars(){R.querySelectorAll(‘.fill’).forEach(function(f){f.style.width=”0″;});setTimeout(function(){R.querySelectorAll(‘.fill’).forEach(function(f){f.style.width=f.dataset.w+’%’;});},40);}
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
