Fish Audio has released Drama 3, a 35-billion parameter mixture-of-experts text-to-speech model that accepts plain-language acting direction and supports 83 languages. The model features open weights available on Hugging Face alongside availability in the Fish Audio API, offering a self-hostable option for developers working on scripted dialogue, audiobooks, and dubbing. Built with a backbone derived from Qwen3.5-35B-A3B and a Codebook-MTP decoder, Drama 3 processes acting cues and bracketed directions to generate 44.1 kHz audio for up to 16 characters in a single request.
On benchmarks, Drama 3 achieved a leading 93.2% overall score on Artificial Analysis Pronunciation Robustness, performing best on brands, places, and technical terms at 95.9%, while scoring an Elo rating of 1242 on the Artificial Analysis quality leaderboard. Running the model requires at least 48 GB of NVIDIA GPU memory in FP8, as CPU, Apple MPS, and AMD ROCm are not supported, with official benchmarking conducted on an NVIDIA H200 GPU using SGLang-Omni serving. The hosted API is priced at $50 per one million characters, with a temporary launch price also available.
Why it matters: Fish Audio Drama 3 expands open-weight access to advanced, directable text-to-speech synthesis with fine-grained emotional control and multi-speaker capabilities.