Fish Audio

Our full take

Fish Audio is a voice generator for produced audio, meaning speech you script, generate and edit before anyone hears it: voiceovers, audiobooks, character and story work. You paste text, pick a voice from a community library the homepage puts at 2,000,000-plus voices or clone your own from a claimed 15-second clip, and export the result. Behind it sits Hanabi AI Inc., a Delaware company, the OpenAudio research line, and the open-source Fish Speech project that made these models some of the most watched in speech AI. What it is not: a phone-agent platform. There are no phone numbers, no SIP trunking (plugging in your own phone carrier) and no per-minute call product anywhere on the site.

Subscriptions are priced in credits, and here are the numbers as the cards print them. Free is $0 for 8,000 credits a month, sized at up to 7 minutes of generated audio. Plus is $15 a month list for 250,000 credits (up to 200 minutes, so about $0.075 a minute). Pro is $100 for 2,000,000 credits (up to 1,620 minutes, about $0.062 a minute). Max is $999 for 25,000,000 credits, and Enterprise is custom volume pricing with zero data retention and on-premise deployment on the card. At capture the page was mid-promotion: a countdown banner priced annual billing at 3 months off plus a holiday 50% off, which is how the cards showed $5.50, $37.50 and $749 a month against struck-through list prices. We record list rates and treat the promotion as promotional.

Now the part that needs clearing up, because the page disagrees with itself. The pricing FAQ says a minute of generation costs roughly 600 to 625 credits. The Plus and Pro cards imply about 1,250 credits a minute (250,000 over 200 minutes; 2,000,000 over 1,620). The Max card implies 4,000 (25,000,000 over 6,250 minutes), which at $999 a month would make the biggest plan cost about $0.16 a minute, more than double the entry tier. No pricing ladder works that way, so something on that card is off, but we record what the page prints and flag it rather than guess. If you are sizing a plan on minutes, get the credits-per-minute figure from Fish Audio in writing first.

Developers get a much cleaner number. The API charges $15 per million UTF-8 bytes on every current model (s1, s2-pro and the recommended s2.1-pro). A byte is effectively a character if you write in English; accented and non-Latin scripts use 2 to 4 bytes per character, so they cost proportionally more. That works out at about $0.015 per 1,000 English characters, and by the docs’ own conversion (a million bytes is roughly 12 hours of speech) about $0.021 per minute of generated audio, the derived figure we carry as this page’s per-minute rate. Against the narration rates we track, that is near the bottom of the table: ElevenLabs stores at $0.10 per 1,000 characters, Cartesia at $0.035, and only Speechify is cheaper at $0.010. Two catches: pay-as-you-go API access needs a paid subscription, and concurrency is earned with spend (5 parallel requests below $100 of lifetime billing, 15 from $100, 50 from $1,000). There is also a free s2.1-pro-free API model for development with fair-use limits, speech-to-text at $0.36 per audio hour, and a Voice Design endpoint at $0.01 a request.

The models are the reason to care. The flagship is s2.1-pro, which the docs credit with 83 languages; s2-pro (80-plus languages) claims roughly 100ms time-to-first-audio on datacentre hardware; the older s1 covers 13 languages with 64-plus emotion tags you write straight into the script in brackets. Note the language claim moves depending on where you read: the homepage still says 30-plus, the docs say 83 for the flagship, and the S2 release post says the training set spans about 50. We store the docs’ flagship figure and flag the spread. The open-source angle is real, not marketing: in March 2026 Fish Audio published the S2 weights, fine-tuning code and a streaming inference engine on GitHub and Hugging Face, and a dated third-party review from June 2026 rates it the leading open-weight text-to-speech family of 2026 on blind preference, while noting ElevenLabs keeps the polish and library advantage. Those are the reviewer’s findings and the vendor’s benchmarks, not ours; we have not run scored listening tests yet.

Could you run a live agent on it? The pieces exist: websocket streaming, that 100ms first-audio claim, a conversational-chatbots solutions page. But there is no telephony, no agent product with a call rate, and the third-party review places real-world latency behind realtime specialists like Cartesia. So on this site Fish Audio registers as a narration option, not a call platform. Wire it into a voice-agent stack as the speech layer if you like the voices; buy a platform if you need the phone part handled for you.

On compliance we record formal claims only. GDPR: yes. The privacy policy carries a proper section for EEA, Swiss and UK users with the full set of data-subject rights and EU standard contractual clauses for transfers. SOC 2: the Enterprise card lists it as a plan feature, but there is no certificate, report or trust page anywhere we could find, so the box stays unticked, the same standard we hold every vendor to. HIPAA: not mentioned, so no. One rights point that catches people: free-plan output is personal use only (the card crosses out commercial use, and the FAQ and the terms agree), and commercial rights on paid plans apply to verified voices you own.

My read: for produced audio the per-character maths is genuinely hard to beat, with open weights as your exit route if the hosted service ever moves against you. Set against that, the plan page’s credit arithmetic needs cleaning up before you commit to a big tier, the compliance paperwork is thin below Enterprise, and none of the certification claims are documented. Buy it for voiceover, audiobook and character work; do not buy it expecting a phone agent. We have not placed test calls or run scored listening tests on Fish Audio, so it carries no Voxrater benchmark numbers and the scores here are an editorial preview. Everything above comes from the vendor’s plan page, docs, terms, privacy policy and blog plus one dated third-party review, captured 2026-07-11, with the pricing screenshot in evidence.