{"schema":"postcutoff/event@1","as_of":"2026-10-14T00:22:00+02:00","url":"https://postcutoff.com/e/2026-10-13-fish-audio-drama-3/","md":"https://postcutoff.com/e/2026-10-13-fish-audio-drama-3/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-13-fish-audio-drama-3","date":"2026-10-13","date_precision":"day","short_title":"Fish Audio releases Drama 3","deck":"A 35B open-weight TTS model directed in plain language, with word-level edits","takeaway":"Weights are on Hugging Face under a new Fish Audio Community License: free for research and non-commercial use, commercial use only below USD 500,000 annual revenue, no hosted services.","category":"model-release","category_label":"Model releases","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Fish Audio: Drama 3 product page","url":"https://fish.audio/drama-3/","type":"official","group":"primary","domain":"fish.audio"},{"n":2,"title":"Hugging Face: fishaudio/drama-3 (model card, weights)","url":"https://huggingface.co/fishaudio/drama-3","type":"code","group":"primary","domain":"huggingface.co"},{"n":3,"title":"Fish Audio Community License v1.0","url":"https://huggingface.co/fishaudio/drama-3/blob/main/LICENSE","type":"official","group":"primary","domain":"huggingface.co"},{"n":4,"title":"Fish Audio Drama 3 Technical Report (PDF)","url":"https://github.com/fishaudio/fish-speech/blob/main/docs/assets/Fish_Audio_Drama_3_Technical_Report.pdf","type":"paper","group":"primary","domain":"github.com"},{"n":5,"title":"GitHub: fishaudio/fish-speech","url":"https://github.com/fishaudio/fish-speech","type":"code","group":"primary","domain":"github.com"},{"n":6,"title":"Fish Audio docs: changelog (Drama 3 Preview, Sept 23)","url":"https://docs.fish.audio/developer-guide/getting-started/changelog","type":"docs","group":"primary","domain":"docs.fish.audio"},{"n":7,"title":"Fish Audio docs source: Drama 3 guide (model headers, tags)","url":"https://github.com/fishaudio/docs/blob/main/features/text-to-speech/drama-3.mdx","type":"docs","group":"primary","domain":"github.com"},{"n":8,"title":"PR Newswire: Fish Audio Launches Drama 3 (Oct 13)","url":"https://www.prnewswire.com/news-releases/fish-audio-launches-drama-3-direct-ai-voices-in-plain-language-control-every-word-and-shift-emotion-mid-sentence-302906620.html","type":"official","group":"primary","domain":"prnewswire.com"},{"n":9,"title":"Fish Audio on X: launch post","url":"https://x.com/FishAudio/status/2110038429199450126","type":"official","group":"primary","domain":"x.com"},{"n":10,"title":"Fish Audio on X: day-one launch partners","url":"https://x.com/FishAudio/status/2110038432856915996","type":"official","group":"primary","domain":"x.com"},{"n":11,"title":"Fish Audio on X: Drama 3 (preview), Sept 23","url":"https://x.com/FishAudio/status/2102805658005635294","type":"official","group":"primary","domain":"x.com"},{"n":12,"title":"OpenRouter: fish-audio/drama-3","url":"https://openrouter.ai/fish-audio/drama-3","type":"docs","group":"primary","domain":"openrouter.ai"},{"n":13,"title":"SGLang-Omni: Drama 3 cookbook","url":"https://sgl-project.github.io/sglang-omni/cookbook/fishaudio_drama3.html","type":"docs","group":"primary","domain":"sgl-project.github.io"},{"n":14,"title":"Unite.AI: Fish Audio Launches Drama 3 to Direct AI Voices and Edit a Single Word","url":"https://www.unite.ai/fish-audio-drama-3-ai-voice-direction-audio-inpainting/","type":"press","group":"press","domain":"unite.ai"},{"n":15,"title":"Artificial Analysis: text-to-speech leaderboard","url":"https://artificialanalysis.ai/text-to-speech/leaderboard","type":"discussion","group":"reactions","domain":"artificialanalysis.ai"}],"official":13,"filed":"2026-10-13","updated":"2026-10-13","orgs":["Fish Audio"],"title":"Fish Audio releases Drama 3: a 35B open-weight TTS model directed in plain language, with word-level edits","summary":"On 2026-10-13 Fish Audio released Drama 3 (preview since 2026-09-23): a text-to-speech model with a 35B-parameter mixture-of-experts backbone (3B active per token) that follows plain-language direction, revises a take from a note, voices multi-character scenes and replaces single words by infilling. Weights are on Hugging Face under a new Fish Audio Community License: free for research and non-commercial use, commercial use only below USD 500,000 annual revenue, no hosted services. \"Most emotive and controllable TTS model in the world\" is the vendor's claim; Artificial Analysis ranks the preview #10 overall (Elo 1243), the top open-weights model, behind Eleven v4.","key_facts":["Architecture (model card, technical report): 35B-parameter MoE text backbone, 3B active per token, hybrid full + linear attention; codebook multi-token-prediction audio decoder with 10 codebooks; keeps the S2 audio tokenizer","Released files on Hugging Face fishaudio/drama-3: 8 safetensors shards (about 39.6 GB, FP8 experts, rest bf16) plus a 1.87 GB codec; inference code in fishaudio/fish-speech and SGLang-Omni","Licence: Fish Audio Community License v1.0 (dated 2026-10-13). Research and non-commercial use free; commercial use only while annual revenue is under USD 500,000; offering a hosted TTS or voice-cloning service or API needs a separate agreement; outputs may not train other speech models. Not an OSI open-source licence","Licence terms on cloning: no imitation of an identifiable person's voice without consent; no removal of 'any watermark or provenance signal applied by the Fish Audio Materials'. The model card and technical report do not describe a watermark","API: POST https://api.fish.audio/v1/tts with model header `drama-3` (released) or `drama-3-preview`; a missing or misspelled header silently falls back to `s2.1-pro` (Fish Audio docs)","Price: Fish Audio's pricing page did not list Drama 3 on 2026-10-13. OpenRouter lists `fish-audio/drama-3` at $2 per 1M UTF-8 bytes. The press release says API usage is 80% off for the first 30 days","Controls (docs): bracketed free-text direction in any language, point tags such as [sighs], paired word tags <whisper> <emphasis> <soft> <fast> <slow> <stress> <forceful>, language tags (experimental), phoneme tags, up to 16 speakers per text","Vendor serving numbers (one NVIDIA H200, SGLang-Omni): time to first audio 88.9 ms p50 at concurrency 1, RTF 0.131; 271 frames/s single stream against 177 for S2","Vendor benchmark (MiniMax Multilingual Testset, 24 languages; ElevenLabs v4 run by Fish Audio through its API): Drama 3 has a lower error rate than ElevenLabs v4 in 9 languages, e.g. English WER 1.13 vs 1.56, and a higher one in others, e.g. Japanese CER 5.02 vs 3.66, Indonesian WER 4.93 vs 1.59","Vendor benchmark (Step-Audio-Edit): style accuracy 76.5 → 81.8 over three revision rounds (Step-Audio-EditX 63.8), but emotion accuracy 58.2 against 69.3 for Step-Audio-EditX","Independent (Artificial Analysis text-to-speech leaderboard, read 2026-10-13): 'Fish Audio Drama 3 Preview' is #10 of 97 models with Elo 1243 (1,857 appearances), the highest-ranked open-weights model, ahead of Breeze TTS 2 (1221) and Fish Audio S2 Pro (1118). Eleven v4 Turbo leads with 1330. The site flags the weights as not commercially open","Languages: about 80 languages and dialects, the coverage of S2 (technical report); the Hugging Face card lists 83 language codes","Company figures in the launch (vendor): 15M+ users, $21M annual recurring revenue, $52M seed round (announced 2026-07-28)","The launch post on X had about 3.24M views after roughly three hours; it offered a free month for commenting 'DRAMA', and its links carry utm_medium=influencer","Demo voice: the product page calls it 'Uncle Roger's licensed voice, directed by Drama 3'. No statement from Nigel Ng (the comedian behind Uncle Roger) was found, and the terms are not public"],"key_numbers":[],"tags":["tts","speech","open-weights","voice-cloning","voice","audio-editing","licence"],"science":null,"body_md":"## What happened\n\nFish Audio released Drama 3 on 2026-10-13, three weeks after a preview (API id `drama-3-preview`, 2026-09-23). The company published the weights on Hugging Face, inference code in `fishaudio/fish-speech`, a technical report, and a press release. The launch post on X said: \"Introducing Drama 3: the most emotive and controllable TTS model in the world. And yes, it's open-weight.\"\n\nThe model is much larger than its predecessor S2 Pro: a 35B-parameter mixture-of-experts backbone that activates 3B parameters per token, against S2's roughly 4.4B. According to the technical report, four functions are new and live in one model:\n\n- **Revision from a note**: the next turn can say \"gentler, and a bit sadder\" and the model redoes the take with the earlier one in context.\n- **Editing by infilling**: one word can be replaced (\"Wednesday\" → \"Thursday\") while the audio around it is kept.\n- **Scene production**: several characters (up to 16 speaker tags) in one pass.\n- **Streaming text input** and long-form synthesis.\n\nDirection is written as free text in square brackets, in any language. Paired tags such as `<whisper>…</whisper>` limit a delivery to specific words, which is what the launch post means by controlling \"how even a single word should sound\".\n\n## Why it matters\n\nDrama 3 puts features that were mostly limited to closed services (take-by-take direction, single-word repair, multi-character scenes) into downloadable weights, and it is the largest open-weights speech model in this dataset. It overtook Breeze TTS 2 as the top open-weights model on Artificial Analysis while staying below the closed Eleven v4 released two weeks earlier. The revenue-capped licence is a middle path between research-only weights and permissive open source.","disputed":[],"related":[{"id":"2026-09-28-elevenlabs-eleven-v4","url":"https://postcutoff.com/e/2026-09-28-elevenlabs-eleven-v4/","date":"2026-09-28","date_precision":"day","short_title":"ElevenLabs launches Eleven v4 and Eleven v4 Turbo, #1 on Artificial Analysis TTS arena","deck":null,"takeaway":null,"category":"model-release","category_label":"Model releases","importance":4,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":11,"official":7,"filed":"2026-09-29","updated":"2026-09-29","orgs":["ElevenLabs"]},{"id":"2026-08-25-breeze-tts-2","url":"https://postcutoff.com/e/2026-08-25-breeze-tts-2/","date":"2026-08-25","date_precision":"day","short_title":"BreezeBlue releases Breeze TTS 2, the new top open-weights text-to-speech model","deck":null,"takeaway":"It pushed the open-weights ceiling in TTS about 90 Elo higher, narrowing the gap to closed leaders (Eleven v4, Cartesia Sonic-3.6).","category":"open-source","category_label":"Open source","importance":3,"confidence":"medium","status":{"key":"partly","labels":["Partly confirmed"]},"sources":4,"official":3,"filed":"2026-09-29","updated":"2026-09-29","orgs":["BreezeBlue"]},{"id":"2026-07-28-fish-audio-52m-seed","url":"https://postcutoff.com/e/2026-07-28-fish-audio-52m-seed/","date":"2026-07-28","date_precision":"day","short_title":"Fish Audio raises a $52M seed round led by Coreline Ventures and Capital Today","deck":null,"takeaway":"$52M is large for a seed round and shows investor interest in voice models outside the biggest labs.","category":"business","category_label":"Business","importance":2,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":3,"official":1,"filed":"2026-10-13","updated":"2026-10-13","orgs":["Fish Audio","Coreline Ventures","Capital Today"]},{"id":"2026-03-09-fish-audio-s2-open-source","url":"https://postcutoff.com/e/2026-03-09-fish-audio-s2-open-source/","date":"2026-03-09","date_precision":"day","short_title":"Fish Audio open-sources S2","deck":"Expressive 80+ language TTS with inline emotion tags","takeaway":"It made open TTS with fine-grained, LLM-style prompt control and production streaming available to the public, and set the open-weights bar for most of 2026.","category":"open-source","category_label":"Open source","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":4,"official":4,"filed":"2026-09-29","updated":"2026-10-13","orgs":["Fish Audio"]}],"people":[],"posts":[{"id":"x-fishaudio-2110038429199450126","title":"Fish Audio introduces Drama 3: 'the most emotive and controllable TTS model in the world', open-weight","url":"https://postcutoff.com/p/x-fishaudio-2110038429199450126/"},{"id":"x-fishaudio-2102805658005635294","title":"“Introducing Drama 3 (preview): the most controllable TTS model ever. Describe the tone…”","url":"https://postcutoff.com/p/x-fishaudio-2102805658005635294/"}],"videos":[{"id":"fish-audio-drama-3-is-here-launch-video","title":"Fish Audio Drama 3 is Here","url":"https://postcutoff.com/v/fish-audio-drama-3-is-here-launch-video/"},{"id":"fish-audio-drama-3-preview-announcement","title":"Fish Audio Drama 3 Is Here: The Most Controllable TTS Model Ever","url":"https://postcutoff.com/v/fish-audio-drama-3-preview-announcement/"},{"id":"comfyui-fish-audio-drama-3-multi-character-dialogue","title":"Fish Audio Drama 3 in ComfyUI: Generate Multi-Character Dialogue Scenes","url":"https://postcutoff.com/v/comfyui-fish-audio-drama-3-multi-character-dialogue/"}],"models":[{"id":"fish-audio-drama-3","name":"Fish Audio Drama 3","url":"https://postcutoff.com/m/fish-audio-drama-3/"}],"changes":[{"date":"2026-10-13","type":"filed","text":"Created"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-13","run":null,"sources_read":null,"updated":"2026-10-13","human_review":null,"version":{"date":"2026-10-13"}},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":166,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":105,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":196,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":135,"in_training_data":false}],"short_url":null}