Fish Audio Releases Drama 3: A 35B Open-Weight MoE Text-to-Speech Model
This week, Fish Audio released Fish Audio Drama 3, an open-weight text-to-speech (TTS) model that accepts plain-language acting direction. At its core, Drama 3 is a 35B mixture-of-experts (MoE) architecture with roughly 3B active parameters per token, offering a self-hostable option for scripted dialogue, audiobooks, and dubbing. The model weights are available on Hugging Face, and the model is also live in the Fish Audio API.
Architecture and Specifications
Drama 3 uses a sparse MoE design with 256 experts, activating 8 per token alongside 1 shared expert. This brings the effective parameter count to approximately 3B at inference, despite the 35B total. The context length has not been disclosed.
Hardware Requirements and Deployment
Drama 3 runs on a single NVIDIA GPU with at least 48 GB of VRAM in FP8 precision. It does not support CPU inference, Apple MPS, or AMD ROCm. Fish Audio's official benchmarks were conducted on a single NVIDIA H200 GPU. This makes the model accessible for local deployment on high-end workstations or cloud instances with a single Hopper-class GPU.
Performance Highlights
On Artificial Analysis, Drama 3 is Fish Audio's highest-scoring model to date. However, it still trails several closed-source APIs on quality Elo. Its strongest area is pronunciation robustness, where it achieves 95.9% accuracy on brands, places, and technical terms in the Artificial Analysis Pronunciation Robustness benchmark. Conversely, its weakest category is exact sequences (codes, paths, emails), scoring 86.3%.
Key Takeaways
- Size: 35B total parameters, ~3B active per token (8 of 256 experts + 1 shared). Context length undisclosed.
- Runs on: 1 NVIDIA GPU with ≥48 GB VRAM in FP8. No CPU, Apple MPS, or AMD ROCm support. Official benchmarks use 1 H200.
- Performance: Fish Audio's top model on Artificial Analysis, yet 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 for directable, multi-speaker speech with open weights. For developers and enterprises, Drama 3 offers a compelling balance of scale, openness, and controllability.
For more details, visit the Fish Audio blog or explore the model on Hugging Face.
via MarkTechPost