AI News Feed
Market watch
Flock CEO Defends Surveillance Technology Amid Backlash, Says ‘We Shouldn’t Pause Progress’Gudea pitches AI 'storm tracker' to spot social media crises before they eruptApple Intelligence Nears China Launch as ASML Expands AI CapacitySeoul Stocks Plunge as AI Doubts and U.S. Inflation Fears Trigger Sidecar, Circuit BreakerUkraine shoots down Russian 'ring' drone after operators initially mistook it for a UFOOpen Specs Could Cut AI Data Center Hardware Integration From a Year to Months
Flock CEO Defends Surveillance Technology Amid Backlash, Says ‘We Shouldn’t Pause Progress’Gudea pitches AI 'storm tracker' to spot social media crises before they eruptApple Intelligence Nears China Launch as ASML Expands AI CapacitySeoul Stocks Plunge as AI Doubts and U.S. Inflation Fears Trigger Sidecar, Circuit BreakerUkraine shoots down Russian 'ring' drone after operators initially mistook it for a UFOOpen Specs Could Cut AI Data Center Hardware Integration From a Year to Months
Large Language Models

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

Fish Audio released Drama 3, a 35B open-weight text-to-speech model that follows plain-language acting directions.

AI News Feed · 2026-10-15 13:46

MarkTechPost reported the release this week. Drama 3 is a mixture-of-experts design with roughly 3 billion parameters active per token, drawn from 8 of 256 experts plus one shared expert. Fish Audio positions it as a self-hostable option for scripted dialogue, audiobooks and dubbing. The context length is not disclosed.

Writers place bracketed cues such as [barely holding it together] before a line to shape delivery, and paired tags change only the words enclosed between them, according to the Drama 3 documentation. Directions can be written in any language independently of the spoken line, and the model supports 83 languages. Speaker tags running from <|speaker:0|> to <|speaker:15|> allow up to 16 characters in one request; the API takes one voice ID per speaker, so a full scene renders in a single call.

The backbone starts from the Qwen3.5-35B-A3B language model and has 40 layers, grouped as three Gated DeltaNet layers per one gated full-attention layer. Only one layer in four keeps a KV cache, while the DeltaNet layers store history in a constant-size state. A new Codebook-MTP decoder replaces the per-frame Fast AR decoder used in S2: four shared layers feed nine layers, one per acoustic codebook, and only codebooks 0 and 1 feed back into the backbone, where S2 fed back all 10. The audio tokenizer is unchanged from S2, built on the Descript Audio Codec, and outputs 44.1 kHz audio at about 21.5 frames per second. Training ran through three pre-training stages, then instruction tuning, then reinforcement learning with GRPO without a KL penalty and a single learned human-preference reward model.

Fish Audio reports Drama 3 at 93.2% overall on Artificial Analysis Pronunciation Robustness, ranked first, against 88.1% for its predecessor S2.1 Pro. Human listeners judged hard English words across four categories: brands, places and technical terms at 95.9%, against 87.3% for S2.1 Pro; context-dependent readings at 95.3%, against 86.1%; shorthand such as numbers, dates and units at 93.6%, against 92.7%; and exact sequences such as codes, paths and emails at 86.3%, against 79.4%. The benchmark tests pronunciation rather than emotional delivery, and Fish Audio describes emotive quality only qualitatively. On the Artificial Analysis quality leaderboard, Drama 3 Preview holds an Elo of 1242, ahead of S2.1 Pro at 1190 but behind Eleven v4 Turbo at 1328.

The Fish Speech install guide lists at least 48 GB of NVIDIA GPU memory in FP8, and states that CPU, Apple MPS and AMD ROCm are not supported. MoE experts ship in FP8, with the remaining weights in bf16. The model card reports streaming results on a single H200: at one stream, median time to first audio is 88.9 ms with a real-time factor of 0.131; at 16 streams, throughput reaches 927.6 frames per second. Serving runs through SGLang-Omni, whose fused kernels are validated only on Hopper GPUs.

In the hosted API, developers set the model header to drama-3 on POST /v1/tts. The documentation warns that a missing or misspelled header silently falls back to s2.1-pro. List price is $50 per 1 million characters, with a $10 launch price running until November 12, 2026.

The weights are released under the Fish Audio Community License v1.0. Research and non-commercial use are free, and commercial use is free below USD 500,000 in annual revenue. Hosting Drama 3 as a third-party API requires a separate license, products must display "Built with Fish Audio," and cloning a real voice without consent is prohibited.

Among close alternatives, Fish Audio's own S2 Pro is a 5B model made of a 4B slow AR component and a 400M fast AR component, released as open weights under a research license, runnable on 24 GB or more of NVIDIA memory or on CPU, with latency of about 100 ms time to first audio on an H200 and an Artificial Analysis quality Elo of 1116, priced at $15 per 1 million characters. ElevenLabs' Eleven v4 Turbo does not disclose its size, lists 90 or more languages, is hosted only, has a median time to first speech of about 150 ms, holds an Elo of 1328 and costs $40 per 1 million characters. Qwen-Audio-3.0-TTS-Plus also does not disclose size or active parameters, lists eight languages plus dialects, is hosted only, holds an Elo of 1266 and costs $20 per 1 million characters at a Singapore endpoint.