Fish Audio review: evidence before a verdict.

What Fish Audio does

Fish Audio offers text-to-speech, voice cloning and streaming through a web app and developer interfaces. Its documentation lists s2.1-pro, s2-pro and s1, while Story Studio supports longer, multi-speaker projects. These are documented capabilities, not findings from our own listening test. Model and capability documentation, checked 15 September 2026.

For a creator, the useful question is whether a voice survives a complete script. For a developer, it is whether the model, limits and billing fit the application. A convincing ten-second sample does not answer either question.

What the hands-on test needs to establish

No paid account, first-hand session notes or vendor exports were supplied for this edition. I have not scored Fish Audio or claimed to have used a particular model. The player demonstration is labelled separately and uses CMU Flite; it is not evidence about Fish Audio.

The test will use the three visible scripts in the listening bench. For each one, I will preserve the first take, request four further takes, and record the exact model, voice ID, settings and generation cost. The final notes will report the number of usable takes rather than showcase only the best result.

Pronunciation and control

The names-and-numbers script includes “Nguyen”, a calendar date and a currency amount. The record must distinguish a correct first take from a corrected take. Corrections need a reproducible setting or text change, with the original and revised recordings kept together. Until that record exists, pronunciation remains unscored.

Emotion and repeatability

The character script turns from relief to urgency. The test will compare whether that change is audible without producing an exaggerated performance. A setting that works once is less useful than one that works across several lines. Any supported expression controls must be documented against the actual model tested.

Developer performance

Latency will be measured separately from voice preference. The session record needs request region, connection, model, input length, time to first audio, total completion time and failures over twenty requests. No latency number in this edition is a measurement from our site.

Pricing and what is not included

The public plan page now displays Plus at US$132 billed annually, equivalent to US$11 per month, with 250,000 monthly credits. This is an annual commitment, not a US$11 month-to-month subscription. The page also shows a US$15 comparator; a monthly checkout quote remains unverified. Fish Audio plans, checked 16 September 2026.

The pricing FAQ’s approximate credit-to-minute conversion does not reconcile with every displayed allowance. Treat the advertised durations as vendor estimates, not guaranteed usable output. Confirm billing mode, taxes, included usage and retry costs in the account. API pricing is separate and uses UTF-8 bytes; see the pricing-unit guide.

Licensing and commercial use

The web terms distinguish personal, non-commercial service use from commercial use under Paid Services, subject to the agreement. A paid plan does not by itself resolve rights to a particular person’s voice or third-party material. Fish Audio terms, rechecked 16 September 2026.

Who should shortlist it

Shortlist it if you can run a small, controlled audition with your own script and check the selected model’s terms. Keep the same finished-audio target when comparing cost; repeated generations belong in that calculation.

Do not buy on the strength of this page if your decision depends on an independent quality verdict. The recording and score are still pending. For a time-critical production, require an audition and confirmed rights before committing to a voice.

The comparison to make

Use the Fish Audio versus ElevenLabs comparison to separate published plan facts from questions that need a recording. For production planning, start with the YouTube voiceover workflow.

Evidence record

Capabilities checked: 15 September 2026. Pricing, model scope and terms rechecked: 16 September 2026. Paid listening session: pending. Model and settings: pending. Account screenshots: pending. First-take audio and failure log: pending. The score will remain blank until those records exist.