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

Fish Audio has released Drama 3, a text-to-speech model built for scripted audio that can follow plain-language acting notes. Writers can drop in cues like [barely holding it together], and the model is meant to turn that into acted speech rather than a flat readout.

The model is a 35B mixture-of-experts system with about 3B active parameters per token. Fish Audio says it supports 83 languages and can handle up to 16 speakers in one request using speaker tags from <|speaker:0|> to <|speaker:15|>. On the hosted API, each speaker gets one voice ID, so a full scene can render in a single call.

Under the hood, Drama 3 starts from Qwen3.5-35B-A3B and uses 40 layers arranged as three Gated DeltaNet layers for every one gated full-attention layer. Only one layer in four keeps a KV cache. The audio stack also changes: Fish Audio replaced the per-frame Fast AR decoder from S2 with a Codebook-MTP decoder, while keeping the same audio tokenizer based on Descript Audio Codec and 44.1 kHz output.

On benchmarks, Fish Audio says Drama 3 reaches 93.2% on Artificial Analysis Pronunciation Robustness, which puts it first there and ahead of S2.1 Pro’s 88.1%. It does especially well on brands, places and technical terms at 95.9%, but exact sequences such as codes, paths and emails are weaker at 86.3%. On the quality leaderboard, the Drama 3 Preview scores 1242 Elo, above S2.1 Pro’s 1190, but below Eleven v4 Turbo’s 1328.

There’s a catch, and it’s not subtle. Self-hosting needs at least 48 GB of NVIDIA GPU memory in FP8, with no support for CPU, Apple MPS or AMD ROCm. Fish Audio also says the weights carry its Community License v1.0, with free commercial use capped at USD 500,000 in annual revenue. The hosted API is live now, and the launch price is $10 until November 12, 2026, before moving to the listed $50 per 1M characters.