Fish Audio has launched Drama 3, a 35-billion parameter open-weight mixture-of-experts (MoE) text-to-speech model designed to generate natural, expressive voice performance from plain-language acting prompts. The model activates roughly 3 billion parameters per token and allows developers to insert direction cues and character speaker tags directly within text scripts across 83 languages.

The model architecture utilizes a Qwen3.5-35B-A3B language model backbone modified with Gated DeltaNet layers to reduce KV cache overhead, paired with a multi-codebook decoder. In evaluation benchmarks, Drama 3 achieved a 93.2% score on the Artificial Analysis Pronunciation Robustness index, ranking first ahead of its predecessor, while recording an Elo score of 1242 on the quality leaderboard.

Model weights have been made available on Hugging Face under the Fish Audio Community License v1.0, which allows free commercial use for entities earning under $500,000 in annual revenue. Hosting on local hardware requires at least 48 GB of VRAM, with optimized inference support available through SGLang-Omni.

Why it matters

  • Brings fine-grained acting direction and multi-speaker dialogue control to open-weight text-to-speech models.

  • Offers a self-hostable MoE voice architecture capable of running on single enterprise GPUs with high throughput.

  • Provides a low-cost API option and commercial license for developers building audiobooks, dubbing, and voice applications.

Source: marktechpost.com