Fish Audio ships Drama 3: open-weight 35B MoE TTS with word-level emotion direction, #1 on Artificial Analysis Pronunciation Robustness at 93.2%

Fish Audio launched Drama 3 on 2026-10-13 as an open-weight TTS model you direct with plain-language bracket tags (for example [barely holding it together]), available in the product, apps, and API (model: drama-3). The backbone is a ~35B MoE with ~3B active parameters per token (top-8 of 256 experts plus a shared expert), initialized from the Qwen3.5-35B-A3B language tower, with hybrid attention (blocks of three Gated DeltaNet layers plus one gated full-attention layer) and a codebook-MTP decoder that predicts ten acoustic codebooks causally while feeding only codebooks 0–1 back into the backbone. Post-training uses a single human-preference reward under GRPO without rule-based ASR/similarity rewards that previously punished laughs and whispers. Independent Artificial Analysis Pronunciation Robustness: 93.2% overall (#1) vs 88.1% for S2.1 Pro, with strong gains on brand/place/tech names (95.9%) and context-dependent readings (95.3%). Vendor H200 streaming: 271 frames/s single stream vs ~177 for S2; concurrency-1 TTFA p50 88.9 ms; concurrency 16 ~927 frames/s at RTF 0.373. Weights: huggingface.co/fishaudio/drama-3 under the Fish Audio Community License (free commercial use under US$500k ARR). Self-host with SGLang-Omni; launch API promo ~$10 per million characters through Nov 12. Technical report: arXiv:2610.18010.