Reviewed August 28, 2026. Kingy ran a zero-spend anonymous demo check and documented the incomplete blind-test method. Paid API use, export fidelity and cancellation were not tested.
Disclosure: Kingy did not purchase credits or a subscription for this update. No affiliate relationship affected the verdict.
The answer in one sentence
The useful finding was not a dramatic “human-like” score. It was the limit itself. One 200-character English test with the Sarah demo voice generated and played in roughly 15.6 seconds. Choosing Adrian reset our text to Fish Audio’s sample script, and a later attempt to regenerate the same controlled script reached the anonymous demo limit and opened a sign-in prompt. No paid credit was used. No download control was visible in that anonymous flow.
That leaves quality, export fidelity and production credit burn unproven. This review publishes the method and empty scorecard rather than inventing results.
Quick verdict
Best for: developers who can test via API, creators with clear voice consent, and personal experiments using Fish Audio’s free model or public demo.
Skip it if: you need a verified blind quality winner today, expect free outputs to carry commercial rights, need a public fixed-duration retention promise, or want to clone a voice without written permission.
Pros
- Official API pricing is simple: the current S2.1 Pro rate is $15 per million UTF-8 bytes, while S2.1 Pro Free is listed at $0.
- The API has no subscription or minimum-spend requirement in the published pricing documentation.
- Fish Audio publishes explicit consent guidance for voice cloning.
Cons
- The anonymous demo did not support a complete controlled three-voice test.
- Selecting another voice reset the entered script, creating an easy test-design error.
- No export button was visible in the tested anonymous demo.
- Commercial rights depend on a paid plan and the current terms.
- Published language counts vary across Fish Audio pages, so model-specific support must be checked.
What Fish Audio is
Fish Audio is a speech platform for text-to-speech, voice design, transcription and voice cloning. Its consumer site offers a web workflow and public examples; its developer documentation exposes models through an API.
Those are not interchangeable products. A free web demo can show whether a voice is interesting, but it does not prove API reliability, batch cost, export metadata or commercial rights. Likewise, an API price does not tell you whether a particular voice pronounces a client’s names correctly.
For context beyond one vendor, compare Kingy’s coverage of the open-source Miso One voice model and Nvidia Fugatto.
How Kingy tested Fish Audio
| Testing dimension | Scope and limit |
|---|---|
| Observed directly | Anonymous homepage demo, voice selection, generation time, script reset, demo-limit dialog and absence of a visible download control |
| Measured directly | One 200-character generation completed in about 15.6 seconds |
| Not measured | Waveform quality, loudness, sample rate, file encoding, API latency, paid credit burn and cancellation |
| Not completed | Controlled three-voice blind listening test |
| Account provenance | No account; no subscription; no API key |
| Cost to Kingy | $0 |
The exact test script was:
At 07:45 on February 29, Dr. Siobhán O’Rourke whispered, “The Łódź prototype costs $1,347.56—really?” Then she laughed, steadied herself, and said: We’ll deliver version 2.1 by Thursday.
It deliberately combines time, dates, a title, Irish and Polish names, currency, decimals, quoted speech, an emotional direction and a version number.
Blind voice test: status and method
Result
Incomplete—no winner declared. The Sarah run completed, but the Adrian selection reset the script and a later controlled attempt hit the anonymous demo limit. Comparing different scripts would confound the result, so Kingy did not score naturalness, pronunciation, emotion or pacing.
The companion files fish-audio-blind-test-script.txt and fish-audio-blind-test-scorecard.csv contain the publishable protocol. To complete it, generate the identical script with three voices under the same model and settings, randomize clips A–C, and collect independent scores from at least three listeners before revealing the voices.
What to score
| Dimension | Question |
|---|---|
| Naturalness | Does it sound continuous and intentional rather than assembled? |
| Pronunciation | Are Siobhán, O’Rourke, Łódź, the currency and version number correct? |
| Emotion | Does the whisper/laugh transition sound plausible rather than theatrical or ignored? |
| Pacing | Are pauses and quoted speech easy to follow? |
| Artifacts | Are there clicks, doubled phonemes, metallic tails or abrupt level changes? |
Raw clips, settings and listener-level scores should be published with the article. Averages without the underlying evidence are not enough.
API pricing and true credit burn
Fish Audio’s pricing and rate-limit documentation listed the following when checked on August 28, 2026:
| Service | Published rate | Important limit |
|---|---|---|
| S2.1 Pro | $15 per million UTF-8 bytes | Byte count is not the same as characters for every language |
| S2.1 Pro Free | $0 | Availability, limits and rights still need model-level checking |
| S2 Pro / S1 | $15 per million UTF-8 bytes | Older-model behavior can differ |
| Speech recognition | $0.36 per audio hour | Input duration drives cost |
| Voice design | $0.01 per successful request | Failed-versus-successful billing should be logged |
Fish Audio estimates one million bytes at roughly 180,000 English words or 12 hours of audio. That is a vendor estimate, not Kingy’s measurement. At the published S2.1 Pro rate, the headline unit cost is attractive; the real project cost still depends on retries, pronunciation repairs, encoding and discarded takes.
A production test should log input bytes, successful and failed requests, returned audio duration and invoice delta. It should report both cost per delivered hour and cost including rejects.
Voice cloning, consent and verification
Fish Audio’s voice-cloning best practices say users should clone only their own voice or one they have permission to use, obtain written permission and avoid public figures without permission. That is the minimum responsible standard, not a paperwork detail.
Fish Audio also describes a professional cloning path with live voiceprint matching and longer reference material. Kingy did not test either instant or professional cloning. Any future review should document who supplied the voice, how consent was captured, whether the reference data can be deleted and how well the system resists a mismatched speaker.
Commercial rights, cancellation and refunds
Fish Audio’s terms distinguish free/personal use from paid commercial use. Users should not assume that a clip generated without payment is cleared for a sponsored video, client deliverable, advertisement or monetized course.
The terms also describe subscriptions as automatically renewing. Cancellation is available through account settings, but fees already paid are generally not prorated or refunded except where the terms or law provide otherwise. Kingy did not purchase, cancel or request a refund, so the ease and timing of that flow remain untested.
Before using an output commercially, save the plan name, generation date, applicable terms and proof of voice consent with the project file.
Privacy and retention
Fish Audio’s privacy policy says content and account data are retained as needed for its systems and legitimate purposes, with deletion rights subject to legal and operational exceptions. It does not give one universal, simple deletion countdown for every kind of content.
Voice data is biometric-adjacent in practical risk even when a policy uses different legal categories. Do not upload a client or employee voice merely because the interface makes cloning easy. Get explicit permission, minimize the recording, avoid unnecessary personal details and confirm deletion controls.
Export fidelity and failure cases
Export fidelity could not be tested because the anonymous demo showed no download control. A complete test should compare WAV and compressed exports, if offered, for duration, sample rate, clipping, leading/trailing silence, metadata and pronunciation. It should also confirm that the downloaded file matches the preview rather than being regenerated.
Observed and credible failure cases include:
- changing voices can reset the script;
- an anonymous demo limit can interrupt a controlled comparison;
- a voice can pronounce ordinary English well but fail names, numbers or code-switching;
- retries can make the real cost higher than the headline API rate;
- a technically good clone may still be unusable because consent or commercial rights are missing;
- Fish Audio pages can describe different language counts, so users should verify the exact model and feature.
Fish Audio FAQ
Is Fish Audio free?
Fish Audio lists a free S2.1 Pro API model and exposes a public demo, but free access does not automatically include commercial rights or unlimited anonymous generations.
How much does Fish Audio cost?
The official API documentation listed S2.1 Pro at $15 per million UTF-8 bytes on August 28, 2026. Consumer subscription pricing and allowances should be checked at checkout.
Can I use Fish Audio commercially?
Paid plans provide commercial use subject to the terms. Free or personal use is non-commercial. Preserve evidence of the plan and terms that applied when you generated the clip.
Can I clone anyone’s voice?
No. Fish Audio’s own guidance requires ownership or permission and warns against cloning celebrities or public figures without permission.
Did Fish Audio win Kingy’s blind test?
No winner was declared. The anonymous limit prevented a fair controlled comparison, and Kingy will not score unmatched scripts.
Final verdict
Fish Audio’s published API economics are compelling enough to justify a controlled trial. Its public demo was fast in one run, but it did not provide enough evidence for a quality verdict, export verdict or true cost-per-delivered-minute calculation. Use Fish Audio for permissioned experiments and API prototyping; delay production dependence until the blind clips, reject-adjusted costs and export files have been audited.
Official sources
- Fish Audio pricing and rate limits
- Fish Audio voice-cloning best practices
- Fish Audio terms
- Fish Audio privacy policy
The Kingy Brief
Get the next Kingy Brief.
Source-checked AI changes, original tests and one practical thing to try.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime