Voice Cloning API: Short Samples to Production Fish Audio
Build a production voice cloning workflow with short samples, consent records, server-side keys, multilingual reuse, testing and credit planning.
A production voice cloning API needs three controls before it generates a single line: a representative authorized sample, a server-side key boundary and a versioned link between the voice identity and the content that uses it. Fish Audio provides cloning from a short sample and exposes the capability through the API included with the $99 Studio pack.
The technical integration is only half the system. Consent, revocation and audit history determine whether the voice can be used safely.
What does a voice cloning API do?
It creates a reusable representation of a speaker from reference audio, then applies that identity to new text. Fish Audio’s voice cloning workflow uses a short recording and keeps the resulting voice available for later generations.
The API is the path for applications that need to create, select or apply voices without asking an operator to use the browser workspace. The current workspace and API overview remain the source of truth for request details and limits.
Which pack includes API access?
Voice cloning is available in every Fish Audio pack, but calling the API requires the $99 Studio pack. Studio includes 15,000 credits and the highest concurrency for real-time streams on this site.
| Use | Pack | API access |
|---|---|---|
| Evaluate cloning in the workspace | $9 Starter | No |
| Commercial projects in the workspace | $29 Creator | No |
| Production API and higher concurrency | $99 Studio | Yes |
Commercial rights start with Creator. API access and credit volume start with Studio. Check pricing before building a production estimate.
How do you capture a useful sample?
Use one speaker, a quiet room and normal speaking pace. Avoid music, second voices, heavy compression and theatrical delivery unless that performance is representative of the project.
Fifteen seconds of clean speech is usually enough. If the project needs a specific accent, character energy or pronunciation, include those traits in the reference. A polished studio recording of the wrong delivery is still the wrong reference.
The consent checklist provides the rights record to capture before upload.
How should the application model a voice?
Store your own stable voice record rather than relying only on a provider display name. At minimum, keep:
- Your internal voice ID.
- Provider voice or model reference.
- Speaker and project owner.
- Source recording hash.
- Consent document and allowed-use scope.
- Creation date, version and retirement status.
- Allowed languages or channel restrictions.
This record lets an application select the correct voice deterministically and makes revocation or replacement possible.
Can a cloned voice work across languages?
Fish Audio supports 30+ languages, and the same voice can carry a script across them. That makes a cloned brand narrator or character consistent across markets.
Localization still needs language review. Pronunciation, register and pacing differ by locale. Test a script containing local names, numbers and a natural interruption before approving a multilingual campaign.
The languages hub links to detailed script guidance.
How do you secure the API?
Keep the Fish Audio key on the server. The browser should authenticate to your application, not to the speech provider.
Generate separate keys for staging and production. Limit access to trusted backend services, rotate keys when staff or contractors change and never place a key in frontend bundles, analytics, build arguments or screenshots.
The text to speech API guide covers retries, streaming and key storage. For concurrent live speech, read the real-time text to speech guide.
What should you test before launch?
Generate a fixed regression script with names, acronyms, currency, two emotions and a sentence break. Compare every new voice version against it.
Listen at normal mobile volume as well as through studio headphones. Measure time to first audio, failure rate and how often human review changes the output.
For interactive products, test interruption. The application must stop playback and cancel unnecessary generation when the user begins speaking again.
How should consent and revocation work?
Record who approved the voice, what they approved and when approval ends. If a speaker revokes permission, stop new generations and follow the agreed plan for existing media.
Do not assume that a commercial pack gives permission to impersonate a speaker. A commercial licence covers platform usage; the speaker’s consent covers the person and project.
What belongs in a launch checklist?
Before production, confirm that the sample is clean, the consent record names the project and the key exists only in the server environment. Test voice selection, generation, timeout handling and a version change in staging.
After launch, monitor how often human review rejects a voice. A rise in corrections can indicate a bad source sample, a new language or a script format the voice does not handle well. Keep the last approved sample and voice version so a regression can be compared against a known baseline instead of memory.
Key takeaways
- Fish Audio clones reusable voices from short authorized samples.
- API access starts with the $99 Studio pack.
- Keep the provider key on your server and version your own voice record.
- Test each target language with local names, numbers and pacing.
- Consent and revocation are part of the system, not paperwork outside it.
Frequently asked questions
How long should a voice cloning sample be?
A clean 15-second recording is enough for many voices. Longer audio can help with unusual voices, but a quiet, representative sample produces better results than a noisy recording.
Does the voice cloning API require the Studio pack?
Yes. API access is included with the $99 Studio pack. Voice cloning itself is available in every Fish Audio pack, but browser-only use does not require API access.
Can a cloned voice speak other languages?
Fish Audio supports 30+ languages and states that a voice is not locked to the language of its original recording. Test names, pronunciation and pacing in each target language before production.
What consent should I collect before cloning?
Document the speaker’s identity, the projects where the voice may be used, the allowed channels, the expiry or revocation terms and confirmation that the recording is authorized.
Do API keys belong in browser code?
No. Keep Fish Audio API keys in a server-side secret manager. A browser should call your backend, and that backend should enforce authentication, authorization and rate limits.