Published Updated
In one or two sentences, before anything else.
Fish Audio is a text-to-speech and voice-cloning platform run by Hanabi AI Inc. of Palo Alto, grown out of the open-source Fish Speech project. In 2026 it removed cloned voices of French dubbing actors after legal notices (April), raised a 52 million USD seed round (July) and released the open-weight Drama 3 model (October 13).
What it can do, one sentence each, and whether we actually tested it.
Reads text in a chosen voice and takes directions written in brackets inside the text, such as [whispering] or [laughing], for emotion, pauses and non-speech sounds.
Per Fish Audio's docs, S2 Pro and S2.1 Pro read bracket cues as ordinary text, so the set of cues is not fixed.
Fish Audio's newest speech model, released October 13, 2026, directed in plain language for tone, pacing and character, with open weights on Hugging Face.
The model card gives 35B parameters with about 3B active per token. API list price 50 USD per million characters, 10 USD until November 12, 2026.
The production API model since June 23, 2026: one model for 83 languages, with multi-speaker dialogue. s2.1-pro-free is the same model at no cost through November 30, 2026.
The free tier has fair-use limits and no latency guarantee, per Fish Audio's docs. S2.1 Pro weights are not released.
Builds a voice from a short reference recording; Fish Audio advises at least 10 seconds of clear speech per clip and says a minute or two improves fidelity.
Fish Audio's docs page on creating voices says nothing about consent; the live ownership check applies to Professional Voice Clones.
Trains a studio-quality clone from 10 to 180 minutes of audio, and finishes only after the speaker reads random text live and the voiceprint matches the uploads.
Launched June 15, 2026; 1, 5 and 15 slots on Plus, Pro and Max. Fish Audio says finished clones cannot be deleted yet.
A library of more than 2,000,000 user-uploaded voices that any user can generate speech with.
Fish Audio's figure. The library is filled by users, which is where disputes over unlicensed clones arise.
Transcribe 1 Pro, released September 18, 2026, transcribes clips up to 60 minutes with speaker turns, emotion cues and optional timestamps.
0.36 USD per audio hour on the API.
The code for running Fish Audio's open models is public on GitHub, so Drama 3, S2 Pro and S1-mini can run on your own GPU.
33.2k GitHub stars on October 17, 2026. Code and Drama 3 are under the Fish Audio Community License; S2 Pro weights are research-only.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
Pairwise blind comparisons of the same text read by two models; scores are Elo ratings from user votes.
What happened
Fish Audio Drama 3 Preview ranks 10th of 97 models with an Elo of 1244 (interval 1229 to 1259), S2.1 Pro 18th (1190) and S2 Pro 35th (1116). ElevenLabs' Eleven v4 Turbo leads at 1330, Cartesia's Sonic 3.6 is 4th at 1285 and MiniMax's Speech 2.8 HD 22nd at 1173. Fish Audio is in the upper group, not at the top.
Case 2
Input
Narration scripts mixing English and Romanian, and a clone of the author's own voice used for a week of YouTube narration.
What happened
He found Fish Audio switched between the languages cleanly while ElevenLabs showed accent bleed and mispronounced some Romanian words, and says viewers did not notice his cloned voice. One person's judgment with no published samples, and the article carries Fish Audio affiliate links.
Case 3
Input
[breathless, speaking fast] Just turn off the lights.
What happened
Fish Audio says Drama 3 reads 93.2% of hard words, numbers and codes correctly on Artificial Analysis's Pronunciation Robustness benchmark, against 88.1% for S2.1 Pro, and advises rendering several takes because they vary. These are the vendor's numbers; Artificial Analysis's public page did not show the scores to us.
Case 4
Input
Upload 10 to 180 minutes of clean recordings of your own voice, then read a random passage live so the system can match it against the uploads.
What happened
Fish Audio says training starts only after the live voiceprint matches, so a clip downloaded from the web should fail. The check covers this tier only, and finished professional clones cannot yet be deleted.
What it does badly, with the evidence, and what we could not settle.
Weak
Fish Audio has hosted clones of real people's voices made without their consent.
Evidence: Eight French dubbing actors served notices dated January 30, 2026 (AFP, Next). By April the group had grown to 25 and Fish Audio had removed 47 models reproducing their voices, but their 20,000 EUR damages claim was not met (ATS via La Télé, KultureGeek with AFP). TechCrunch reported on July 28, 2026 that takedowns had been slow and that a voice stays up until its owner files for removal.
The free plan is for personal, non-commercial use only.
Evidence: The terms (effective August 18, 2024) and the plan page FAQ limit free users to personal, non-commercial projects; commercial use needs a paid plan, even though the free plan card lists commercial use.
The open weights are not free for every business.
Evidence: S2 Pro weights are under the Fish Audio Research License, which grants no commercial rights. Drama 3 and the code use the Fish Audio Community License: commercial use only below 500,000 USD annual revenue, and never as a hosted speech or voice-cloning service without a separate licence.
Your uploads and usage can train Fish Audio's models.
Evidence: The terms let Fish Audio use usage data and content to develop and train its AI models, and public uploads carry a perpetual licence that can outlast account deletion.
Who should not use it now: brands, studios and agencies that need voices with cleared rights, and anyone planning to clone a voice they do not own.
Evidence: The public library is user-uploaded, the French case shows unlicensed clones reaching the platform, and the Community License ends automatically if you imitate an identifiable person without consent. ElevenLabs and Cartesia also rank higher in the blind arena.
Uncertain
Which company holds the model rights.
Why unsure: The terms and site footer name Hanabi AI Inc., while both model licences name 39 AI, Inc. as copyright holder; Fish Audio does not explain the relationship.
Whether quick clones get any ownership check.
Why unsure: Fish Audio describes a live voiceprint check only for Professional Voice Clones; its developer page on creating voices says nothing about consent.
The growth figures.
Why unsure: The 21 million USD in annual recurring revenue and 8 million users come from Fish Audio's July 2026 press release and are not independently audited.
How the French dispute ends.
Why unsure: Damages were unresolved in April 2026, and the actors' lawyer said he was preparing a new case to ban such platforms in France; we found no ruling.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Maker | Hanabi AI Inc., a Delaware corporation; headquarters in Palo Alto, California | Headquarters per Fish Audio's July 2026 press release. Co-founders Shijia Liao (chief scientist, formerly at Nvidia) and Rissa Cao (chief executive), per TechCrunch. Source: fish.audiochecked |
| Funding | 52 million USD seed round announced July 28, 2026, led by Coreline Ventures and Capital Today | The release claims 21 million USD in annual recurring revenue and more than 8 million users. Source: aijourn.comchecked |
| Plans and prices | Free 0 USD; Plus 15 USD a month or 132 USD a year; Pro 100 USD a month or 900 USD a year; Max 999 USD a month or 8,988 USD a year; Enterprise custom | Source: fish.audiochecked |
| Free plan | 8,000 credits a month (up to about 7 minutes), 500 characters per generation, 3 public voice slots; personal, non-commercial use only | The FAQ and the terms restrict free users to non-commercial use. Source: fish.audiochecked |
| API prices | S2.1 Pro and S2 Pro 15 USD per million UTF-8 bytes; s2.1-pro-free 0 USD through November 30, 2026; speech to text 0.36 USD per audio hour | A Chinese character takes 3 bytes in UTF-8, so Chinese text costs more per character than English. Source: docs.fish.audiochecked |
| Drama 3 API price | 50 USD per million characters; 10 USD until November 12, 2026 | Source: fish.audiochecked |
| Open weight licences | Drama 3: Fish Audio Community License v1.0 (commercial use below 500,000 USD revenue); S2 Pro: Fish Audio Research License (non-commercial); S1-mini: CC BY-NC-SA 4.0 | S1-mini licence from its Hugging Face card; the Drama 3 card's metadata still carries a research-licence tag. Source: fish.audiochecked |
| Artificial Analysis Speech Arena | Drama 3 Preview 10th (Elo 1244), S2.1 Pro 18th (1190), S2 Pro 35th (1116), OpenAudio S1 47th (1081), of 97 models | Source: artificialanalysis.aichecked |
| Apps and platforms | Web app, Fish Audio app for iOS and Android, API | Source: fish.audiochecked |
| Model retirement | S1 is deprecated and retires December 31, 2026; s1 requests then run on S2.1 Pro at the same price | Source: docs.fish.audiochecked |
Answer first, then why, an example, the limits and what to do next.
Fish Audio is a US-registered company: its terms name Hanabi AI Inc., a Delaware corporation, and its press release puts headquarters in Palo Alto. Co-founder and chief scientist Shijia Liao, a former Nvidia researcher, started it as the open-source Fish Speech project, and Rissa Cao is chief executive. Its July 2026 seed round was co-led by Coreline Ventures and the Chinese firm Capital Today.
Fish Audio has a free plan with 8,000 credits a month, about 7 minutes of speech, but only for personal, non-commercial use. Commercial use starts with Plus at 15 USD a month or 132 USD a year. Developers can also call S2.1 Pro through the API at no cost as s2.1-pro-free until November 30, 2026, within fair-use limits.
Drama 3 is free for commercial use only if you and your affiliates earn under 500,000 USD a year, under the Fish Audio Community License v1.0. Above that, or to run it as a hosted speech or voice-cloning service, you need a licence from Fish Audio. S2 Pro weights are for research and non-commercial use only, and S1-mini is CC BY-NC-SA 4.0.
Yes, French dubbing actors found clones of their voices on Fish Audio, and Fish Audio removed 47 such models by April 2026. Eight actors, including Angelina Jolie's French voice, served notices dated January 30, 2026 demanding removal and 20,000 EUR in damages (AFP). By April the group numbered 25; damages were not paid, and their lawyer said he was preparing a new case.
Not in blind listening tests: ElevenLabs' v4 models hold the top two places on the Artificial Analysis Speech Arena (Elo 1330 and 1329), while Fish Audio's best, Drama 3 Preview, is 10th at 1244. Fish Audio is far cheaper on the API, 15 USD per million UTF-8 bytes for S2.1 Pro against 80 USD per million characters for Eleven v4, and it publishes open weights.
ElevenLabs, Cartesia, MiniMax Speech and CosyVoice are the closest alternatives, each for a different need. ElevenLabs and Cartesia's Sonic 3.6 (4th, Elo 1285) rank above Fish Audio in the blind arena; MiniMax's Speech 2.8 HD is 22nd (1173), below Fish Audio's S2.1 Pro. CosyVoice is an open-source model family with Apache-2.0 code, nine languages and 18 or more Chinese dialects; it is not on the arena.
This page is built from Fish Audio's own pages, docs and model licences, its July 2026 funding release, news reports, the Artificial Analysis Speech Arena and one hands-on comparison by Bitdoze. Every fact links to its source, and we checked each one on October 17, 2026.
Fish Audio is run by Hanabi AI Inc., a Delaware corporation (terms), with headquarters in Palo Alto (press release). Co-founder and chief scientist Shijia Liao, a former Nvidia researcher, began it as Fish Speech, a voice model trained on a single GPU and open-sourced; Rissa Cao is chief executive (TechCrunch). PEdaily counts a team of 22 (PEdaily).
| Date | Event |
|---|---|
| December 23, 2023 | First Fish Speech release in the changelog, v0.2.0 (changelog) |
| June 2025 | S1 (4B) and the open S1-mini (0.5B) |
| February 3, 2026 | Eight French dubbing actors' notices to Fish Audio and VoiceDub made public (AFP via France 24) |
| March 2026 | S2 with open weights (changelog) |
| June 15, 2026 | Professional Voice Clone with a live ownership check (Fish Audio) |
| June 23, 2026 | S2.1 Pro, offered only as a service |
| July 28, 2026 | 52 million USD seed round; the company claims 21 million USD in annual recurring revenue and 8 million users (press release) |
| October 13, 2026 | Drama 3, with open weights (Fish Audio) |
The Fish Speech repository had 33.2k GitHub stars when we checked (GitHub), and Fish Audio says its public library holds more than 2,000,000 user-uploaded voices (Fish Audio).
| Model | Released | Weights | Licence |
|---|---|---|---|
| Drama 3 (35B, about 3B active) | October 13, 2026 | Open | Community License: commercial use below 500,000 USD revenue, no hosted service |
| S2.1 Pro | June 23, 2026 | Not released | Service only |
| S2 Pro | March 2026 | Open | Research License: non-commercial only |
| S1-mini (0.5B) | June 2025 | Open | CC BY-NC-SA 4.0 |
Sources: Drama 3 announcement, Community License, Research License, S1-mini card. S1 retires on December 31, 2026 (pricing docs).
| Plan | Monthly | Yearly | Credits a month |
|---|---|---|---|
| Free | 0 USD | None | 8,000 (up to about 7 minutes) |
| Plus | 15 USD | 132 USD | 250,000 |
| Pro | 100 USD | 900 USD | 2,000,000 |
| Max | 999 USD | 8,988 USD | 25,000,000 |
Source: Fish Audio plan page. On the API, S2.1 Pro and S2 Pro cost 15 USD per million UTF-8 bytes, and s2.1-pro-free runs the same model at no cost until November 30, 2026 (pricing docs). Drama 3 lists at 50 USD per million characters, cut to 10 USD until November 12 (Fish Audio).
Fish Audio's terms forbid content that infringes other people's rights, the Community License bans imitating an identifiable person without consent, and Professional Voice Clones are trained only after the speaker passes a live voiceprint check (guide). The public library, though, is filled by users. On January 30, 2026, eight French dubbing actors, among them Angelina Jolie's French voice Françoise Cadol, gave Fish Audio and VoiceDub eight days to remove clones of their voices and claimed 20,000 EUR in damages (AFP via France 24, Next). By April their lawyer Jonathan Elkaim said Fish Audio had removed 47 models for 25 actors but paid no damages, and that he was preparing a new case (La Télé, KultureGeek). TechCrunch reported in July that takedowns had been slow; Fish Audio says they now take under three minutes once an owner proves the voice is theirs (TechCrunch).
| Best Speech Arena rank | API price | |
|---|---|---|
| Fish Audio | 10th, Drama 3 Preview (1244) | 15 USD per million UTF-8 bytes (S2.1 Pro) |
| ElevenLabs | 1st, Eleven v4 Turbo (1330) | 80 USD per million characters for v4, 40 USD for v4 Turbo (ElevenLabs) |
| Cartesia | 4th, Sonic 3.6 (1285) | Not checked |
| MiniMax | 22nd, Speech 2.8 HD (1173) | Not checked |
| CosyVoice | Not listed | Free to self-host |
Of these, Fish Audio and CosyVoice publish models you can run yourself; CosyVoice's code is Apache-2.0 and covers nine languages and 18 or more Chinese dialects (GitHub).
Sources
Changelog
VideoEnhancer is a web tool that upscales short MP4 clips to 1080p, 2K or 4K with the FlashVSR model and also compresses, face-swaps and generates video. After a 5-second free trial, 1080p costs 3 credits a second, $0.03 to $0.09 depending on how credits are bought. VideoEnhancer is our own product: it is made by the team behind GenGrades.
LayerGrab splits a flat image into up to 16 named, transparent layers over a filled-in background, on the web, in Figma or in Chrome, or pulls out only what you name. Each layer costs $0.10 on its $12 plan, $0.08 on Premium. No independent review exists yet. LayerGrab is our own product: it is made by the team behind GenGrades.
VideoTextRemover is a web tool that erases burned-in subtitles, captions and other on-screen text from MP4 and MOV videos and repaints what was behind them, up to 1080p. It bills per second processed, about $0.025 to $0.049, with 30 free seconds a month in some countries. VideoTextRemover is our own product: it is made by the team behind GenGrades.
MagicRemover is a browser tool that erases people, objects, text, logos and watermarks from photos, and subtitles from videos. Photos get 2 free removals a day, 2 more after sign-in, at about 1 megapixel; beyond that, credits start at $9.90 for 330. MagicRemover is our own product: it is made by the team behind GenGrades.
Hermes Agent is the best OpenClaw alternative for most people: free, open source, safer by default and able to import an OpenClaw setup in one command. NanoClaw puts every session in a container, ZeroClaw is a tiny sandboxed binary, Claude Cowork suits people who would rather pay than host, and Hostinger is the cheapest way to keep OpenClaw itself.
EditTextImage is a web tool, with phone apps and a Chrome extension, that rewrites, fixes, translates or removes the words in a finished image while keeping its lettering. Credits are sold once, from $14.90 for 40; a standard edit costs one. No independent test exists yet. EditTextImage is our own product: it is made by the team behind GenGrades.