Fish Audio

By Minel Gunesoglu, founder. Nothing on this page describes how Fish Audio sounds: no audio was generated and no voice was cloned for it, and the quality figure below belongs to somebody else's blind-listening board. What I did do was read the vendor's Terms of Use end to end, read its pricing page in a browser with both billing views open, pull its own Voice Cloning FAQ, and set a year of customer reports from four platforms against those documents. Everything here was read on 12 August 2026, apart from the vendor's own voice-library count, which was captured a day later, and all of it is re-checked monthly.

One of the sites that answers to this product's name says in its own footer that it is not this product. So the first job of a Fish Audio review is smaller and more useful than a verdict: work out which thing you are looking at. Fish Audio is the hosted text-to-speech and voice-cloning platform at fish.audio, and its own Terms of Use name Hanabi AI Inc. as the entity providing the service. That platform, and nothing else, is what the score on this page measures.

Three other things share the name. Fish Speech, now branded OpenAudio, is the same company's open-weight model line; it is a genuinely different product with a different licence, and it gets one bounded section near the end of this page rather than being folded into the numbers. Big Fish Audio is an unrelated company that has sold sample libraries and loops since 1986, and it surfaces in the same search results purely on the shared word. Then there are the look-alikes: fishaudio.org carries this product's branding throughout while stating in its own footer that it is independently operated by a company called Kitta AI and is not the official Fish Audio service. That footer is in Chinese and was captured to bytes on 13 August 2026; the English above is this desk's rendering of it, not a quotation. None of those domains is linked anywhere on this page, and none of them is a source for any statement here about Fish Audio's product, pricing or terms, the one thing any of them is cited for is what its own footer says about itself.

Fish Audio Trust and Identity: Is fish.audio the Real Site?

fish.audio is the address to check against, and the confirmation that holds up is a document rather than a badge. The Terms of Use served at that domain print their own date, "Effective date: August 18, 2024", and name the operator directly: "The entity providing the Services is Hanabi AI Inc., a corporation operating under the laws of Delaware". The footer of the same site's Voice Cloning page carries the same company name. Both were read on 12 August 2026.

Where the company actually sits is a question this review does not answer. Published accounts disagree, this desk has read three different answers to it and verified none of them at source, so the only location claim on this page is the one printed in the contract itself.

Is Fish Audio Safe? What the Dated Record Shows

Safe is the searcher's word, and it is not a word this page will hand back as a verdict. What can be shown is a dated record, and the customer half of it is small. Trustpilot files the profile under the name Fish, with fish.audio as the linked site, and on 12 August 2026 it read 2.5 out of 5 across 23 reviews, 22 of them posted inside the previous twelve months. Trustpilot's own label for that band is Poor. The profile is unclaimed and carries no vendor reply anywhere on it.

Read by subject rather than by star rating, those 23 reviews point somewhere specific. The negative ones are overwhelmingly about money and about reaching support: refunds, invoicing, credit consumption, a cancel path reviewers say they could not find. Voice quality is a minority complaint, which matters because it means the low rating and the capability score below are measuring different things rather than contradicting each other.

Two further sources are worth reporting as absences, and they are not the same kind of absence. Hacker News is a genuine dated zero: its public search API returns no hits at all for fish.audio, fish-speech or the plain two-word name, so no launch thread or discussion has ever registered there. Reddit is weaker than that. Two independent sweeps surfaced no on-topic thread about this product, which is reported here as unconfirmed rather than as proof that none exists.

Fish Audio's Official App and Download, and the Look-Alikes to Avoid

There is no official app-store badge on Fish Audio's own homepage to anchor a reader, checked on 12 August 2026, and that absence is what gives the look-alikes their room. One site, fishaudio.org, resolves itself: its footer brands the site Kitta Audio, copyrights it to Kitta AI, carries the line "Powered by Fish Audio", and states (in Chinese, read and captured on 13 August 2026), that it is independently operated by Kitta AI and is not an official Fish Audio service. An App Store listing published under an individual developer's name points at a fourth domain, fishaudio.online. Two further bare domains, fishaudio.cloud and fishaudio.app, and two further App Store listings were checked by nobody and are treated here as unverified rather than as either official or fake.

The practical rule that falls out of this is short. Type the address rather than following a result, confirm it reads fish.audio, and treat any other spelling of the name as a different company until its own pages say otherwise. I read six dedicated Fish Audio reviews for this page on 12 August 2026, running from roughly 1,650 to roughly 6,500 words, and found this question raised in not one of them.

Fish Audio Pricing and the Free Tier: What $0 Actually Gets You

The number you see for Fish Audio's entry paid tier depends on which view of the pricing page happens to load. Read in a browser in US dollars on 12 August 2026, with the billing control toggled by hand, the Plus tier renders three ways in one afternoon: $15 a month struck through, $7.50 a month under a banner reading "Anniversary: first month 50% OFF", and $5.50 a month against a line reading "$66 billed annually". The struck-through $15 is identical in both billing views, and the vendor's own arithmetic on the annual card confirms what it is: a saving of $114 measured against twelve months at $15. So $15 is the standard monthly rate, $7.50 is a first-month promotion, and $5.50 sits under two stacked discounts.

Currency is a second layer on the same control. Read from this desk the page defaults to euros and to the annual view, which is why every figure below was taken in dollars in both views instead. The euro figures the same page serves are not conversions of the dollar ones, so anyone outside the United States should read their own currency off the page rather than converting these.

One more figure belongs beside the price list because buyers keep colliding with it. Fish Audio's own pricing FAQ prices generation at roughly 600 to 625 credits a minute. A Trustpilot reviewer dated 2 May 2026 describes a very different rate: "In just 30 seconds of audio generation, it can consume around 2000 credits, which feels like an unfair and aggressive pricing model." Seven months earlier, on 14 October 2025, an unconnected reviewer reported a different mismatch on the same theme: "It claims to handle up to 30,000 characters, but it only processed about 926 characters from a short 50-second audio clip." Neither account can be verified from a review page, and neither is treated here as a measurement. What is on the record is that the vendor publishes a rate, two strangers seven months apart report being metered at something else, and the page reconciles none of it.

Fish Audio Free Plan Limits: 500 Characters Per Generation, and the Commercial-Use Row Resolved

Fish Audio's free tier carries two separate limits, and the one that stops people first is not the headline. The monthly pool is 8,000 credits, advertised as up to 7 minutes of generation, alongside 3 public voice slots. The other limit sits on every single request: "Up to 500 characters per generation". Paste one page of a script and you hit that wall immediately, with the monthly pool barely touched. The paid step above it moves the same cap to 15,000 characters.

The commercial-use question on that tier had a contradiction attached to it, and the contradiction is now closed. Records on this desk carried the vendor's pricing page as saying free-tier commercial use was allowed while the Terms of Use said the opposite. Read in a browser on 12 August 2026, the Free column's Commercial use row is drawn with a line through it and a muted icon: it is that column's excluded row, sitting beside Enhanced voice cloning, which is struck the same way. The disagreement was an artefact of reading extracted text, which cannot carry a strikethrough. The same page's FAQ says it in prose too: "Free plan users can only use generated content for personal, non-commercial projects." The free tier is a place to find out whether the voices suit the work, and not a place to make anything you intend to sell.

Fish Audio Plus, Pro and Max Plans: What Each Tier Costs and Unlocks

Every figure in this table was read off Fish Audio's own pricing page on 12 August 2026, in US dollars, in both billing views.

The commercial-use column needs its source named, because the plan cards do not carry it as text. On the Free Tier card the "Commercial use" row is drawn with a line through it and a muted icon rather than a tick, and a text extraction of that page renders it as an ordinary included feature. The Plus card writes it out as "Commercial use allowed". Above Plus the row is not restated on the card at all, so those cells rest on the vendor's own FAQ on the same page: "Premium subscribers can use verified voices (that you own) for commercial purposes. Free plan users can only use generated content for personal, non-commercial projects." Read in both billing views, 12 August 2026.

Plus is the tier that matters most, because it is where commercial use begins and therefore where the ladder actually starts for anyone earning from the output. It adds unlimited public voice slots plus 10 private ones and one professional voice slot. Pro adds three team seats and five professional slots; Max adds ten seats and fifteen. Nothing above Plus changes the rights position, so the money above it buys capacity and seats rather than a better licence.

One row moves between views, and it is the row a cautious buyer would most want to be stable. "7 days money back guarantee" appears in the Pro column of the annual view. It appears nowhere in the monthly view, and it is absent from the Plus column in both. A buyer on the entry paid tier is not covered by it in either billing state, which is the fact that keeps the refund position below where it is.

Fish Audio API Pricing: Billed Per UTF-8 Byte, Not Per Character

Fish Audio's pricing page answers the API question by pointing elsewhere: "Yes! Premium subscribers get access to our flexible pay-as-you-go API." The developer documentation it points at, read on 12 August 2026, prices generation per UTF-8 byte rather than per character, and that unit is the whole finding.

A UTF-8 byte is not a character. In plain English text the two run close enough to ignore, because unaccented Latin letters take one byte each. Accented European text, Cyrillic, Arabic, Japanese and Chinese characters and emoji take two to four bytes each, so a script in those languages bills at two to four times what a per-character estimate predicts. Any comparison table that converts this rate into a price per million characters is quietly using the wrong unit for most of the world's writing systems.

The rate itself is deliberately not reprinted here. That documentation page was read but never captured to bytes on this desk, and a price this page cannot reopen and re-check is a price it will not publish. Read it on the vendor's own developer documentation before you budget an integration.

Fish Audio Voice Cloning: Consent Rules and the Character Voice Library

Cloning is the product Fish Audio leads with, and the vendor describes the mechanic plainly: a short reference clip, a voice available in seconds, and zero-shot output across 13 languages without re-recording. The reference length is stated two ways on that one page. Its headline reads "Clone any voice in 15 seconds.", while the feature list and the FAQ below it both say ten seconds is enough, the FAQ in the words "Ten seconds of clean speech is enough." Both were read from the same capture on 12 August 2026. It is a small discrepancy with a practical edge: it is exactly the kind of number a third-party comparison table copies from whichever line it happened to read.

Language coverage past the core set has a visible record, and it sits on the open-source side of the house rather than on the hosted platform. The fish-speech repository is busy: 32,161 stars, 2,757 forks, 689 issues opened over its life and 18 open on 12 August 2026. That activity is what makes two quiet threads stand out. A Vietnamese pronunciation report filed on 3 June 2026 was specific, saying "It seems to be the only language underperforming—I've tested 50+ voices and personal clones, but the pronunciation is consistently off." No human replied to it. Its only two comments came from the repository's own stale-issue bot, the second of which closed it on 19 July 2026: "This issue was closed because it has been inactive for 14 days since being marked as stale." Two Punjabi reports filed on 10 August 2026, one of them noting that "The Gurmukhi version produces a noticeably different pronunciation from the Latin versions.", were still open and unanswered two days later. None of that shows those languages perform badly. It shows that when a specific-language gap was reported, this desk could find no path from the report to a human answer.

Fish Audio Voice Cloning Consent and Rights: What the FAQ and Terms of Use Say

Fish Audio does not check whose voice you are cloning, and it says so itself. Its Voice Cloning FAQ places the obligation on the uploader: "You are responsible for confirming you have the rights, consents, and disclosures required for any voice you clone", and the same answer states that "Fish Audio does not pre-clear individual use cases". Higher up the same page, the marketing copy invites the use that rule is about, opening an example list with "Have a sitting president narrate your dating app". Both sentences were read from one capture of one page on 12 August 2026, and they are reported here as they sit, without any claim about what the company meant by publishing them together.

The complaint that answers that policy is in French, and it is invisible unless you ask for it. Trustpilot's default view of this profile shows English only; switching to all languages surfaces five French reviews posted between February and July 2026. One of them, dated 16 July 2026, reads: "Des voleurs qui s'approprient le travail de professionnels sans leur accord, honteux." Rendered into English by this desk: the reviewer is calling the company thieves who appropriate professionals' work without their consent, and calling that shameful. Those are the reviewer's words and the reviewer's characterisation, not a finding of this desk, and no individual allegation on a review page can be verified from the review page. What is verifiable is narrower and still worth a professional's attention: the vendor's own stated policy and the platform's loudest non-English complaint describe the same mechanism from opposite sides of it.

Two boundaries belong with that. This desk has not surveyed how other voice platforms handle the same question, so nothing here says whether Fish Audio's posture is common or unusual. And in many places consent for a named person's voice is required by law regardless of what any platform checks, which is a question for a lawyer rather than for a review. One clause in the Terms of Use is worth reading before you upload anything either way: "Usage Data and Content may be used to develop, train, or enhance artificial intelligence or machine learning models".

Fish Audio's Free Character Voice Library: 2,000,000+ Community Voices

Fish Audio's homepage advertises more than 2,000,000 community-uploaded voices, captured to bytes on 13 August 2026. The count appears three times on that one page, most plainly as "The Fish Audio platform hosts over 2,000,000 voices". It is the vendor's own figure and nobody has audited it. It also moves, and the comparison that follows is ours rather than the vendor's: this desk's launch-queue note from five weeks earlier recorded the claim at 200,000-plus, our note, not a capture of the page on that date, which is a weaker thing and worth saying. Taken together they describe a marketing number rather than a specification, and the size of a voice library says nothing about how many of its voices are any good.

The library also inherits the consent rule above, which is the part a commercial user should carry. Those voices were uploaded by other users, and confirming rights and consent is the uploader's stated responsibility rather than something the platform clears in advance. A voice you find in the library therefore comes with whatever rights checking its uploader did, and Fish Audio's free tier gives you 3 public voice slots to publish your own into the same pool.

How Fish Audio's 71.0 Score Was Built: All Four Dimensions Measured

The interesting thing about 71.0 is not the total but which end of the card the pull comes from. Capability is the highest row here and the contract is the lowest, and the two were read on the same day from two sources that have nothing to do with each other: a public blind-listening leaderboard and a legal document.

An arena Elo is worth one plain sentence, because it is the least familiar number on the card. Listeners on Artificial Analysis' Speech Arena hear the same text read by two systems without being told which is which, they vote, and the rating rises or falls with the result, exactly as a chess rating would. Fish Audio S2.1 Pro sits at 1,141 on 1,826 of those votes, which is a solid upper-middle placement in a field of 93 rather than a contender's one. The board publishes its own uncertainty alongside the placing, and it puts this entry's true position somewhere between 12th and 21st. It also moves under you: this desk has watched the same row shift within a single week, which is why the figure above carries its read date and why the capture behind it carries an expiry.

The composite is a single average across the cells that carry a value, and on this card that is now all four of them. It was three until 12 August 2026, when the usability cell filled and the total moved from 58.4 to 62.6. Both of those are that day's figures, since overtaken by changes to the scale itself rather than to anything measured with it. A card with three filled cells and a card with four are not the same kind of number, which is why the count sits beside the score rather than behind it. The rules every card here is built under are public, including that one.

Two things stay outside the arithmetic on purpose. The first is what people report about the voices, which is where the record turns strange: on 14 March 2026 one Trustpilot reviewer wrote "Ran blind listening tests with my team. Fish Audio S1 won against ElevenLabs 7 out of 10 times." and, on that same day and the same profile, another wrote "Totally sucks. Wasted a lot of time and some money. Weird stuff happens with the voices." Two customers rated the same company on one day and landed nowhere near each other, and it is not simply an old-model-against-new split, because both were writing after the S1 generation shipped. A third, dated 2 May 2026, reported that "The voice quality sounds robotic and unnatural, making it unsuitable for professional or even casual use." Nothing in that set is a measurement, and none of it feeds a cell. The second exclusion is Fish Audio's own published comparison of its models against named competitors. That is a vendor's own material about other vendors, and this desk does not use that class of evidence on anybody's card, so it appears nowhere in the number above.

Where Fish Audio's Usability Score Comes From, and the Blank This Page Withdrew

This cell shipped blank, and the blank has been withdrawn. It rested on a finding that Product Hunt, which files this product under the slug fish-speech, refused every route the community-mining pass had, and a refusal to a fetcher is not the same thing as a page that cannot be read. Read again in a rendered browser on 12 August 2026, that page returned HTTP 200 with its full review text, the Cloudflare interstitial clearing on its own with no challenge presented and none solved. The correction is about this desk's method rather than about Fish Audio, which is why it is printed here rather than applied quietly.

Product Hunt publishes 4.7 across 12 reviews for this product, read and captured to bytes on 12 August 2026. That is not the 76.9 on the card, and the distance between the two is deliberate. A 4.7 out of 5 is 94.0 on this site's hundred-point scale, and twelve reviews is a thin sample, so the model pulls the figure toward the middle of the scale in proportion to how little of it there is, which lands the cell at 76.9. Twelve is just above the minimum sample this site will score a review aggregate on at all, so the cell is measured at the confidence twelve reviews can carry rather than at face value.

Three other reads belong here, because they say what this cell does not rest on. G2 answered with a CAPTCHA; the product's own slug was confirmed to exist from a search-result title, but no review content was ever reachable. Capterra's category pages returned HTTP 403 on each attempt, so no Fish Audio profile there was ever confirmed to exist and none was guessed at. Hacker News is the different case and the honest one: its public search API answers, and the answer is zero results. Two refusals and one genuine zero, and the number above rests on Product Hunt alone.

Trustpilot's 2.5 does not stand behind the cell either, and it never could have. Ease of use and consumer satisfaction with billing and support are different constructs, and the complaint mix behind that rating is mostly the second one. A 2.5 on one platform and a 4.7 on another are two measurements of two different things rather than a contradiction to be averaged away.

That has been this page's argument since 12 August 2026, and until 18 August the card was quietly doing the opposite: the engine pooled Trustpilot's 23 reviews with Product Hunt's 12 by head count and published 67.4, in the same table as a section explaining why 76.9 was the honest figure. Since 18 August the card does what the page says. The row reads 76.9, the composite 71.0, and the 2.5 across 23 reviews keeps the job it always did here: dated prose in the sections above, with the billing complaints behind it named one at a time.

Two things about that deserve suspicion rather than trust. The row went up while the evidence beneath it got smaller, from 35 readings to 12, and twelve is barely above the minimum this site will convert a review aggregate on at all; the card marks the cell thin, which is the only honest label for it. And the two records are held apart on how each sample was gathered, not on what either says: every review visible on that Trustpilot profile carries the platform's unprompted label, and how Product Hunt's twelve arrived has never been read. If it turns out they arrived the same way, the two pool again and this row comes back down.

Fish Audio's Commercial Terms: the Contract Behind the 56.3

Fish Audio's Terms of Use were read in full on 12 August 2026, and the document's own effective date is 18 August 2024, so it was already close to two years old on the day it was read. Four positions in it produce the 56.3, and the most consequential of them is an absence: across 58,426 characters of vendor text, the word Output does not appear once.

That absence matters because the document is otherwise generous about permission and silent about ownership. Paid users are granted a commercial-use licence in plain terms, "if you are a user of Paid Services, you are licensed to use the Services for commercial uses", and the document then never says who holds rights in the audio the model produces. The nearest ownership language points at other things: "This is a license only – your ownership in User Submissions is not affected." concerns what you upload rather than what the model generates, and "Using the Services does not give you ownership of or rights to any aspect of the Services" runs the other direction entirely. Fish Audio's marketing surfaces answer more warmly than its contract does. The Voice Cloning FAQ says "Yes. Paid plans include commercial rights", and the pricing page's Plus column reads "Commercial use allowed". Those are marketing pages; this dimension reads legal documents, and on the legal document the paid-tier output position is recorded as unclear after a full read rather than as a grant.

The free tier is the one position the contract states without ambiguity: "You will only use the Services for your own internal, personal, non-commercial use". That is the same rule the pricing page's struck-through row carries, arrived at from the opposite end of the site.

Fish Audio's Indemnification Clause: Who Indemnifies Whom

The indemnity runs one way, from you to the company. The clause reads "You agree to indemnify and hold the Fish.Audio Parties harmless from and against any and all claims, liabilities, damages", and it attaches to your use of the service, including anything a third party does with your account, and to any breach of the agreement by you. Every occurrence of that root across the whole document runs the same direction; no clause anywhere runs from the vendor to the user, and no tier is carved out either way.

This desk's own rubric treats a one-way indemnity as the industry default rather than as an outlier, which is why the position scores at the middle of its ladder rather than at the bottom, and why it still carries the heaviest weight of the four. For a solo creator posting to their own channel it is an abstract risk. For anybody putting a synthetic voice into a client deliverable it is the sentence to read before signing, because it is the contract's answer to who carries a third-party claim about the finished file, and the answer is the buyer.

Fish Audio's Refund Policy and the Cancellation Complaints Behind It

The refund position is the plainest thing in the document, set in capitals: "FISH.AUDIO WILL NOT REFUND ANY FEES THAT YOU HAVE ALREADY PAID". The sentence before it removes the partial case as well, stating that on cancellation "you will not be eligible for a prorated refund of any portion of the subscription fee paid for the then-current subscription period". Whether a statutory escape hatch sits behind that was checked rather than assumed: the document contains no money-back, cooling-off, withdrawal or consumer-law language of any kind, and its one reference to a statutory right is the jury-trial waiver. The seven-day guarantee named earlier is the only refund path in the documents this desk read (the Terms of Use, the pricing page in both billing views, and the Voice Cloning FAQ, all on 12 August 2026) and it is scoped to one tier in one billing view.

The customer record on this exact point converges from two platforms that have nothing to do with each other, which is what makes it worth reading closely. On 5 July 2025 a Hugging Face discussion opened on Fish Audio's own model repository with the line "I signed up for the monthly plan, they billed YEARLY and there is no way to end the subscription". Fifty-two seconds later the thread was closed by an account that HuggingFace's own data marks as a member of the Fish Audio organisation. It kept collecting replies from strangers describing the same problem for the next seven weeks, and on 25 August 2025 a second org-affiliated account answered once: "After you make a payment, there’s always a ‘Manage Subscription’ button on your plan page." A customer replied the same day that the button he found offered no way to cancel. Both accounts sit in the same thread and this desk cannot referee between them; the disagreement is the finding and it is left open here.

Six months later the same complaint shape appears somewhere else entirely. Four Trustpilot reviews between January and April 2026 describe a cancel path that is unreachable and refund requests answered after weeks or not at all. The earliest, dated 5 January 2026, reads: "They would not let me cancel and hid the cancel option." The vendor's own contract, read against those accounts, says the cancellation route is in the account settings: "To change or resign your Paid Services at any time, go to account settings." What a buyer can take from the pair is not a verdict about the company but a sequencing rule. Because fees already paid are not returned, the decision to make carefully with Fish Audio is the first charge, not the cancellation.

Fish Speech and OpenAudio: Fish Audio's Open-Source Model and Its Commercial License

Fish Speech and OpenAudio are one thing under two names, and saying so once saves a developer an hour. The repository and the Hugging Face organisation are still named fish-speech; the current model series built on that foundation is branded OpenAudio, as in OpenAudio S1 and S1-mini. Same team, same line, two labels in circulation.

The licence is where the assumption usually breaks. The fish-speech repository describes itself as SOTA Open Source TTS, and its licence field reads FISH AUDIO RESEARCH LICENSE, which is a custom licence rather than one of the standard permissive ones. Downloading and self-hosting the weights is free for research and personal work. Deploying them commercially requires a separate paid licence agreed with Fish Audio directly, and paying for the hosted API or a Plus plan does not satisfy it. The citable technical description of the model family is the team's own paper, arXiv 2411.01156.

One scoping note keeps this section from contaminating the card above it. Nothing in this review's score describes the open-weight line. The capability row reads the hosted flagship, S2.1 Pro, at rank 16, and that row carries no open-weight tag on the board (which is the whole reason the scoping note exists. Four Fish Audio entries sit on that same board. The tagged one is S2 Pro at rank 23, Elo 1,123 across 1,883 votes, and the board says so in its own words: "Fish Audio S2 Pro is the highest-ranked open weights model on the Text to Speech Leaderboard with an Elo score of 1123." Two more sit lower), OpenAudio S1 at rank 41 with an Elo of 1,072 across 4,415 votes, and OpenAudio S1 Mini at rank 57 on 1,046. Scoring the tagged row instead of the flagship would have cost seven places and 18 Elo points, which is why the model name beside a leaderboard figure matters as much as the figure.

Fish Audio Alternatives and the Verdict: Is It Worth It?

There is one cross-tool fact on this page and this is it. On the same leaderboard and the same read date of 12 August 2026, ElevenLabs' Eleven v3 sits at rank 11 of 93 with an Elo of 1,171, against Fish Audio S2.1 Pro's 1,141 at rank 16, and the board's own uncertainty bands for the two do not overlap. ElevenLabs carries its own card here, built on the same four dimensions and read on the same day, and the voice-generators category page orders every scored tool by its composite at build time. This site does not stage products against each other beyond a dated fact like that one.

So, worth it for whom. Fish Audio fits somebody who wants credit-metered speech cheaply, who is testing on the free tier with no intention of selling what comes out of it, and who understands that the entry paid tier at $15 a month is where commercial use starts. The community library and the ten-second cloning mechanic are real conveniences, the arena placement says the current flagship is a competent model rather than a novelty, and the price of finding that out for yourself is $0 within the 500-character limit.

Three readers should not buy it, and each has a document behind the advice rather than an opinion. Anyone delivering voice to a client under a contract that requires clear ownership of the finished audio should not sign this one, because the word Output does not appear in it and the nearest ownership language is about something else. Anyone who needs a refund route should assume there is not one: fees already paid are not returned, and the seven-day guarantee is scoped to the Pro tier in the annual view alone. And anyone planning to clone a voice they do not own should read the consent rule first, because the platform states that it does not pre-clear individual use cases and the obligation lands on the uploader.

The last limit is this page's own. Nobody here has heard a Fish Audio track. The quality figure is a third party's blind-listening board, the price is the vendor's page read in both views, the terms are the vendor's document read end to end, and the customer record is 23 Trustpilot reviews, 12 Product Hunt reviews, one Hugging Face thread and a GitHub tracker. That is enough to settle what you would be buying and what you would be signing. It is not enough to tell you what it sounds like, and that question is left open here rather than answered out of other people's adjectives.

Minel Gunesoglu built this score card and wrote the page. Connect on LinkedIn, or read who builds these cards. Every rating, clause and price above carries the date it was read and is re-checked monthly. Disclosure: some outbound links on this site may be affiliate links; a paid submission only ever buys a faster review, never a listing and never a better score.