What is Fish Audio Drama 3?
Fish Audio Drama 3 is a text-to-speech and voice-cloning model from Fish Audio, the trading name of 39 AI, Inc., whose technical report went up on 13 October 2026. It is a mixture-of-experts model with 35 billion parameters, about 3 billion active per token, with the experts stored in FP8 and the rest in bf16. One script can carry up to 16 speaker tags, and the model clones voices from one or more reference clips, follows written direction such as [sighs] or [laughs], edits a word inside an existing take, and streams long text chunk by chunk. Fish Audio says it was trained on about 80 languages and dialects; its report evaluates 24.
The weights are on Hugging Face under the Fish Audio Community License 1.0, not an open-source licence: research and non-commercial use are free, commercial use is free only while you and your affiliates earn under $500,000 a year, and offering it as a hosted voice service always needs a separate deal. It runs on SGLang-Omni and vLLM-Omni from launch.
On Fish Audio's own API (model drama-3) the list price is $50 per million UTF-8 bytes of input text, discounted to $10 until 17:00 UTC on 12 November 2026. Fish Audio puts one million bytes at about 12 hours of English speech.
How much does Fish Audio Drama 3 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Community License; commercial use under $500,000 annual revenue; no hosted services |
| API list price | $50 per 1M UTF-8 bytes | |
| API launch price | $10 per 1M UTF-8 bytes | Until 17:00 UTC, 12 November 2026 |
Weights free under the Fish Audio Community License for research and for businesses under $500,000 annual revenue. API: $50 per 1M UTF-8 bytes, $10 until 12 November 2026.
Is Fish Audio Drama 3 worth it?
Drama 3 is the open-weight voice model to beat. On Artificial Analysis's blind-vote leaderboard it holds an Elo of 1,243, tenth of all models and first among open weights, ahead of BreezeBlue's Breeze TTS 2 at 1,228 though still 92 points behind ElevenLabs' Eleven v4. It also leads Artificial Analysis's pronunciation test at 93.2%, ahead of Eleven v4's 91.7%, which matters more than naturalness for anyone reading product names, codes or addresses aloud.
Cloning is its weaker side. On the same site's cloned-voice board it ranks 17th at 1,065, and Fish Audio's own report shows the gap: on LibriSpeech-PC it posts the worst word error rate (1.85) and speaker similarity (0.605) in a table that includes OmniVoice and Qwen3-TTS, figures it quotes from other papers. Its multilingual results are uneven too, self-reported: English error of 1.13 against 1.56 for ElevenLabs v4, but 5.02 in Japanese, 6.35 in Hindi and 22.70 in Cantonese. The report contains no human listening tests of its own.
The licence decides who can use it. Commercial use is free only below $500,000 in annual revenue across you and your affiliates, with a 90-day transition after crossing it, and any service whose main purpose is letting outsiders generate speech or clone voices needs a separate licence whatever your size. Shipped products must carry 'Built with Fish Audio', and outputs may not be used to train another speech model.
Consent is a licence term, not a feature. The licence forbids imitating an identifiable person without their prior consent (a parent's for a minor, the rights holder's for someone who has died) and requires you to bind your own users to the same rules, with the licence ending automatically on breach. Neither the model card nor the technical report describes a consent check, a speaker verification step or an audio watermark, so the safeguard is whatever your product builds.
Serving is reasonable for the size. On one H200, Fish Audio measures first audio in 88.9 ms for a single stream and a real-time factor of 0.373 at 16 concurrent streams, so one GPU keeps about 16 voices talking faster than they play. The catch for small teams is that the checkpoint needs a data-centre card, not a laptop.
Where it earns its place
An audiobook studio with eight staff producing dramatised fiction
Voice a full cast from one script without booking actors for minor parts
Tag each character with a speaker tag (up to 16 per text) and direct delivery inline. The licence allows a studio to sell finished audio to publishers as long as clients get no self-serve access, and at about 12 hours of speech per million bytes the API costs roughly $4 an hour of audio at list price.
An ML engineer at an indie game studio under $500,000 revenue
Generate and patch NPC dialogue on an in-house GPU
Serve the weights with SGLang-Omni on one H200, which Fish Audio measures at a real-time factor of 0.373 across 16 streams. When a writer changes a line, the previous take goes into previous_takes and only the edited words are regenerated.
A product lead at a voice-cloning start-up
Decide whether the open weights can power a public cloning feature
They cannot without a separate licence: a service whose main purpose is letting users clone voices counts as a Hosted Service regardless of revenue. Fish Audio's own API is the licensed route, at $50 per million bytes after 12 November and 5 to 50 concurrent requests depending on prepaid spend.
Before you commit
- Setup
- Install SGLang-Omni with the fishaudio extra and run sgl-omni serve --model-path fishaudio/drama-3, or use vLLM-Omni with vllm serve --omni; each request takes text plus one or more reference clips with transcripts.
- Integrations
- SGLang-Omni and vLLM-Omni serve it with streaming; Fish Audio's API exposes it as drama-3 alongside S2.1 Pro, and the fish.audio web app uses the same credits as S2.1 Pro.
- Data and privacy
- Fish Audio Community License 1.0: free for research and non-commercial use, commercial use only under $500,000 combined annual revenue, no hosted voice services without a written licence, 'Built with Fish Audio' attribution, no impersonation without consent, no training other speech models on outputs. Self-hosted audio stays on your hardware.
- Cost at scale
- The safetensors hold 32.2 billion FP8 and 3.7 billion bf16 parameters, roughly 40 GB of weights by our arithmetic, so plan on one 80 GB or larger GPU; Fish Audio's figures are on an H200. On the API, 100 hours of English audio is about 8.3 million bytes: around $83 at the $10 launch price and $417 at the $50 list price, and Chinese, Japanese and Korean text uses about three bytes per character.
- Limits
- API concurrency is 5 requests under $100 prepaid, 15 from $100 and 50 from $1,000, shared across keys and speech-to-text, with HTTP 429 and no Retry-After header beyond it. Self-hosted, time to first audio rises to 450.9 ms at 64 streams on one H200.
What's good
- Elo 1,243 on Artificial Analysis, first among open-weight models
- 93.2% on Artificial Analysis pronunciation robustness, ahead of Eleven v4
- Up to 16 speakers per script, inline direction tags and word-level editing
- 88.9 ms to first audio on one H200 (self-reported)
What's not
- Free commercial use stops at $500,000 annual revenue; hosted voice services always need a deal
- 17th on Artificial Analysis cloned-voice board
- Weak in Japanese, Hindi and Cantonese on its own report
- No built-in consent check or watermark described
Best for
Small studios, game developers and app makers under the revenue cap who want near-ElevenLabs narration and multi-character dialogue on their own GPUs.
Fish Audio Drama 3 alternatives
Fish Audio Drama 3: common questions
- Is Fish Audio Drama 3 worth it?
- Drama 3 gives open weights a voice listeners rank alongside the paid leaders, under a licence that stops being free the moment a business or a voice API gets serious. We rate it 4.2 out of 5.
- How much does Fish Audio Drama 3 cost?
- Weights free under the Fish Audio Community License for research and for businesses under $500,000 annual revenue. API: $50 per 1M UTF-8 bytes, $10 until 12 November 2026.
- What are the downsides of Fish Audio Drama 3?
- Free commercial use stops at $500,000 annual revenue; hosted voice services always need a deal; 17th on Artificial Analysis cloned-voice board; weak in Japanese, Hindi and Cantonese on its own report; no built-in consent check or watermark described.
- Who is Fish Audio Drama 3 best for?
- Small studios, game developers and app makers under the revenue cap who want near-ElevenLabs narration and multi-character dialogue on their own GPUs.
- How much does Fish Audio Drama 3 cost for a team or in production?
- The safetensors hold 32.2 billion FP8 and 3.7 billion bf16 parameters, roughly 40 GB of weights by our arithmetic, so plan on one 80 GB or larger GPU; Fish Audio's figures are on an H200. On the API, 100 hours of English audio is about 8.3 million bytes: around $83 at the $10 launch price and $417 at the $50 list price, and Chinese, Japanese and Korean text uses about three bytes per character.
- Is Fish Audio Drama 3 safe for company data?
- Fish Audio Community License 1.0: free for research and non-commercial use, commercial use only under $500,000 combined annual revenue, no hosted voice services without a written licence, 'Built with Fish Audio' attribution, no impersonation without consent, no training other speech models on outputs. Self-hosted audio stays on your hardware.
- What are the best Fish Audio Drama 3 alternatives?
- The closest alternatives we have reviewed are Eleven v4 and v4 Turbo, Breeze TTS 2, OmniVoice, Qwen3-TTS and Cartesia Sonic 3.6.
Answers checked Oct 28, 2026.
Make Fish Audio Drama 3? Show this rating on your site
