Fish Audio Review

Open Weights, Briefly

• Fish Audio is a voice AI platform providing text-to-speech and voice cloning, with earlier model generations released as open source you can run yourself.
• S2.1 Pro is the current model, described by the vendor as ultra low latency and leading on expressive control.
• Cloning needs roughly ten to fifteen seconds of audio, which is fast enough to be genuinely convenient and fast enough to be a concern.
• A community marketplace holds over two million voices, uploaded by users — the largest library in the category and the part that needs the most scrutiny.
• 30+ languages, REST APIs and SDKs with streaming, and a free monthly allowance before usage-based or subscription plans.

Models, Coverage And Access

What Open Weights Actually Buy You Here

Fish Audio open-sources earlier generations of its models, and that decision does more for a certain kind of buyer than any quality claim.

You can run it yourself. For anyone whose audio cannot leave their infrastructure — health, legal, defence, or simply a client contract that forbids third-party processing — self-hosting is not a preference. It is the only configuration that is permitted at all.

You can audit it. A security team can inspect what the model does rather than accepting a description, which is frequently the condition for approval in a regulated environment.

You are not exposed to someone else's decisions. Voice AI has consolidated fast and prices have moved. A model you hold is insurance against a repricing, a deprecation or an acquisition.

Cost changes shape. Hosted APIs charge per character or per second, so a high-volume application pays continuously. Self-hosting converts that into a fixed infrastructure cost, which flips the arithmetic entirely above a certain volume.

The trade-offs are real and worth stating. The open models are earlier generations, so you are not self-hosting the current flagship. You take on GPU costs, deployment and maintenance. And support becomes a GitHub issue rather than a contract. But the option existing at all is what distinguishes this from ElevenLabs and the other closed providers.

A hosted voice API is a rate you accept. An open model is a rate you can walk away from — which is the only leverage anyone has in a consolidating market.

Two Million Voices Is Not An Unambiguous Good

The community marketplace is the headline number and it deserves careful reading rather than admiration.

The genuine upside is variety. Two million voices means you will find something close to what you want without commissioning a recording, across accents, ages, registers and languages that a curated library would never cover.

The question the site does not address is where those voices came from. A community marketplace of uploaded voices, on a platform where cloning takes ten seconds of audio, invites an obvious concern: how many of those two million are clones of people who did not consent?

That is not an accusation about this platform specifically — it is the structural problem with every open voice marketplace, and it is the reason it deserves stating plainly. Ten seconds of audio is a podcast clip, a YouTube video, a voicemail.

The practical implication for a commercial user is concrete rather than philosophical. If you build a product on a community voice and it later transpires that voice belongs to an identifiable person who never agreed, the exposure is yours as much as the platform's. A voice is a person's likeness, protected independently of copyright in several jurisdictions, with new legislation arriving specifically for synthetic voices.

The defensible approach is narrow: use your own cloned voice, use a voice you have written permission for, or use a voice the platform can confirm the provenance of. For anything commercial, ask that question before building rather than after.

Ten-Second Cloning, Both Ways

Cloning from ten to fifteen seconds of audio is a genuine technical achievement and it is worth thinking about what it means in both directions.

The legitimate uses are strong. A creator can clone their own voice and narrate a back catalogue without recording sessions. An accessibility application can preserve someone's voice before a medical procedure affects their speech. A company can maintain a consistent brand voice across a content library produced over years by different people.

The other direction is equally obvious and does not require malice to cause harm. Ten seconds is a trivial amount of publicly available audio for almost anyone who has ever appeared on a podcast, a webinar, a video call recording or social media.

This is why the consent question above is not academic, and why any organisation deploying voice cloning should have a written position on it before the first clone rather than after the first incident.

For your own voice, the questions to settle are what happens to the model when you close the account, whether it can generate audio you have not authorised, and whether self-hosting lets you keep the model rather than leaving it on someone else's platform. That last one is a genuine advantage of the open-weight route.

None of this is a reason to avoid the technology. It is the diligence the technology requires, and it applies equally to every vendor in the category.

1. Test the current hosted model and an open-weight release on the same script — the gap tells you what self-hosting costs in quality.
2. Measure latency from your own infrastructure rather than trusting a published figure.
3. Clone your own voice first and judge the fidelity on speech you know intimately.
4. Before using any marketplace voice commercially, ask the platform about its provenance.
5. Write down your organisation's consent position before you clone anyone else's voice.

Free Allowance, Then Usage

This site does not publish plan prices. Fish Audio's structure has three distinct routes and the differences matter more than the numbers.

There is a free monthly generation allowance for personal use, usage-based and subscription plans that carry commercial rights, and self-hosting of the open models as a third path entirely.

The most important thing to notice is that commercial rights are tied to the paid tiers. That is a common structure and an easy one to fall foul of — generating audio on a free allowance and shipping it in a product is the kind of mistake discovered late and awkwardly. Establish which tier permits commercial release before you build anything.

Usage-based charging is the right shape for infrastructure and it has the standard consequence: cost scales with success. For a consumer application where users generate freely, model cost per active user carefully, because voice generation is not proportional to what a user pays you.

The self-hosting route changes this completely. Above a certain volume the fixed cost of running your own inference undercuts per-character API pricing substantially, and that crossover point is worth calculating rather than assuming. It is the main reason to care that the models are open at all.

One question specific to the marketplace: whether using a community voice carries different commercial terms from using your own clone. Those are not obviously the same thing and the answer should be in writing.

Where It Sits Against The Alternatives

Placing this accurately is more useful than praising it, and the category is unusually well differentiated.

Against ElevenLabs: ElevenLabs remains the quality and ecosystem benchmark, with more products built around the models. Fish Audio's counter is openness, the marketplace and self-hosting. If quality is the only axis, ElevenLabs; if control matters, Fish.

Against Inworld: Inworld is infrastructure obsessed with realtime latency and model routing for conversational agents. Fish Audio is closer to a voice platform with an API. They overlap on TTS and differ on what surrounds it.

Against Murf: Murf is a finished voiceover product for people who want to make a narration, not build a system. Different buyer entirely.

Against ACE Studio: singing synthesis is a genuinely different problem, and ACE Studio is where that lives.

The honest summary is that Fish Audio is the option for people who want to keep their options — developers building on voice who do not want a single vendor holding the model, the rate card and the roadmap.

• Choose Fish for: self-hosting, auditability and marketplace variety
• Choose ElevenLabs for: peak voice quality and the surrounding product ecosystem
• Choose Inworld for: realtime conversational latency and model routing
• Choose Murf for: a finished voiceover product rather than an API
• Choose ACE Studio for: singing, which is a different problem entirely

Voice AI Terms Used Above

Four terms that decide what you are actually comparing between vendors.

Open weights

The trained model parameters are published, so you can run, inspect and modify the model yourself. Distinct from open training data or open source code.

Voice cloning

Building a synthetic voice from a sample of real speech. The sample requirement has fallen from minutes to seconds, which is what makes consent the live issue.

Zero-shot cloning

Producing a usable voice from a very short sample without retraining a model. What makes ten-second cloning possible.

Streaming TTS

Audio returned as it is generated rather than after the whole utterance completes. Essential for conversational applications, irrelevant for narration.