Models

Fish Audio Debuts Drama 3 Open-Weight Speech Model

Fish Audio has released Drama 3, an open-weight 35-billion parameter text-to-speech model that gives developers high-level control over emotional acting and multi-speaker dialogue.

MarkTechPost2 days agoModels
Image: MarkTechPost

Fish Audio has launched Drama 3, a 35-billion parameter mixture-of-experts text-to-speech model with roughly 3 billion active parameters per token. Built on the Qwen3.5-35B-A3B backbone, the architecture uses 40 layers grouped into three Gated DeltaNet layers for every gated full-attention layer, with only one in four layers retaining a KV cache. A new Codebook-MTP decoder feeds back codebooks 0 and 1 into the backbone, replacing the older S2 model decoder. For self-hosting, the model requires an NVIDIA GPU with at least 48 GB of VRAM in FP8 precision, excluding support for CPU, Apple MPS, or AMD ROCm. Evaluated on a single NVIDIA H200 using SGLang-Omni, streaming tests yielded a median time to first audio of 88.9 ms with a real-time factor of 0.131 at one stream, scaling to 927.6 frames per second across 16 streams.

Targeting audiobooks and dubbing, Drama 3 processes natural-language acting directions, accepting emotional tags in any language across 83 spoken languages. It allows up to 16 unique speaker tags per request to synthesize multi-character scenes in a single API call. Audio is generated via the Descript Audio Codec at 44.1 kHz and roughly 21.5 frames per second. Training involved three pre-training stages, instruction tuning, and reinforcement learning via GRPO without a KL penalty using a human-preference reward model.

On the Artificial Analysis Pronunciation Robustness benchmark, Drama 3 scored 93.2% overall, topping predecessor S2.1 Pro's 88.1%. It recorded 95.9% on technical terms, 95.3% on context readings, 93.6% on shorthand, and 86.3% on exact sequences like codes and emails. On the quality leaderboard, Drama 3 Preview registered an Elo of 1242, placing it ahead of S2.1 Pro's 1190 and S2 Pro's 1116, but behind Eleven v4 Turbo at 1328 and Qwen-Audio-3.0-TTS-Plus at 1266. The hosted API costs $50 per 1 million characters, reduced to a $10 launch price until November 12, 2026. Under the Fish Audio Community License v1.0, open weights are free for research and commercial entities earning under $500,000 annually.

This is our own summary of reporting by MarkTechPost

More in Models