Drama 3
Fish Audio's open-weight 35B MoE text-to-speech model with bracketed stage directions across 83 languages.
Overview
Fish Audio released Drama 3 on October 13, 2026, a 35B-parameter mixture-of-experts text-to-speech model with about 3B active parameters per token. Users direct delivery with bracketed instructions for tone, emotion, or sounds like a breath or laugh, and paired tags such as <emphasis> affect only the wrapped words, in any of 83 languages. It scores 93.2% on Artificial Analysis' Pronunciation Robustness benchmark, ranking first against 88.1% for S2.1 Pro. Interactive editing and audio inpainting are in invite-only testing, with an editing API planned.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 35B MoE (about 3B active per token), 44.1 kHz audio tokenizer, 271 frames/s decode on one H200
- Bracketed stage directions and paired tags for emotion, emphasis, and non-speech sounds
- 83 languages; directions can be written in any language
- 93.2% on Artificial Analysis Pronunciation Robustness (S2.1 Pro: 88.1%)
- Open weights under the Fish Audio Community License v1.0; voice cloning without consent prohibited
- Available on fish.audio, the Fish Audio app, and the API (model: drama-3)
Use cases
- Audiobook and character voice work with directed performance
- Multi-speaker dialogue and game or video voiceover
- Voice agents needing accurate brand, place, and technical-term pronunciation
Pricing
API list price $50 per million characters; $10 per million characters (80% off) from Oct 13 to Nov 12, 2026. Weights free for research and non-commercial use, and for commercial use under $500K annual revenue.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 13 Oct 2026
- Record updated
- 18 Oct 2026
- Record last checked
- 18 Oct 2026