Fish Audio Releases Drama 3, a 35B Open-Weight Text-to-Speech Model

Fish Audio released Drama 3 this week, introducing a new text-to-speech model designed for scripted dialogue and audiobooks. According to MarkTechPost, the 35B mixture-of-experts model features approximately 3B active parameters per token and is available via the Fish Audio API and as open weights on Hugging Face.

The model accepts plain-language acting directions to generate acted speech, allowing writers to insert cues like [barely holding it together] before lines. Drama 3 supports 83 languages and enables up to 16 characters per request through speaker tags <|speaker:0|> to <|speaker:15|>. The API requires one voice ID per speaker, rendering full scenes in a single call.

Technical Architecture and Performance

Drama 3’s backbone originates from the Qwen3.5-35B-A3B language model, featuring 40 layers grouped as three Gated DeltaNet layers per gated full-attention layer. Only one layer in four retains a KV cache, while DeltaNet layers store history in a constant-size state. A new Codebook-MTP decoder replaces the per-frame Fast AR decoder from the previous S2 model, using four shared layers feeding nine layers, one per acoustic codebook.

Performance benchmarks from Artificial Analysis show Drama 3 as Fish Audio’s highest-scoring model, though it trails several closed APIs in quality Elo. It achieved 95.9% accuracy on brands, places, and technical terms, but scored 86.3% on exact sequences like codes and emails, its weakest category. The audio tokenizer remains unchanged from S2, built on the Descript Audio Codec to output 44.1 kHz audio at about 21.5 frames per second.

Hardware and Licensing Constraints

Running Drama 3 requires specific hardware, as it needs one NVIDIA GPU with at least 48 GB of VRAM to operate in FP8 precision. The model does not support CPUs, Apple MPS, or AMD ROCm. Official benchmarks were conducted using a single H200 GPU. The open weights are distributed under a revenue-capped community license, which may limit commercial use cases for developers seeking unrestricted deployment options.

Training involved three pre-training stages, followed by instruction tuning and reinforcement learning. The reinforcement learning phase utilized GRPO without a KL penalty and a single learned human-preference reward model. This approach aims to improve the model’s ability to follow complex acting directions and maintain consistent voice characteristics across long-form content.

What should you do?

Security and technology professionals should evaluate the model’s licensing terms before integrating it into commercial applications, as the revenue cap may restrict certain use cases. Developers interested in self-hosting should verify their hardware meets the 48 GB NVIDIA GPU requirement, as no alternative hardware support is currently available. Teams should also test pronunciation accuracy for technical terms and codes, given the model’s lower performance in those specific categories.

What to do and how to stay safe: Fish Audio

• Review the revenue-capped community license to ensure compliance with your organization’s commercial usage policies.
• Verify that your infrastructure includes NVIDIA GPUs with at least 48 GB of VRAM, as no other hardware is supported.
• Conduct pronunciation tests on technical terms, codes, and email addresses to assess suitability for your specific content needs.
• Monitor the Hugging Face repository and Fish API for updates on context length and potential hardware support expansions.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What hardware is required to run Drama 3?

Drama 3 requires at least one NVIDIA GPU with 48 GB of VRAM to run in FP8 precision, with no support for CPUs, Apple MPS, or AMD ROCm.

How many languages does Drama 3 support?

The model supports 83 languages and allows up to 16 speakers per request using specific speaker tags.

Is Drama 3 available as open weights?

Yes, the weights are available on Hugging Face under a revenue-capped community license, and the model is also live in the Fish Audio API.