Niches · Voice cloning & custom voices
Turn a recording you hold into a reusable voice
No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.
What counts as done
Docs state the shortest and longest sample accepted, file size and formats, whether a transcript or language is required, and what preprocessing (denoise, enhancement, speaker separation) runs; one API call (or one call plus a documented status poll) returns a voice id usable in TTS; a clip outside the limits is refused with a named error rather than accepted and failing later; and a voice made through the API matches one made through the vendor's web app from the same clip.
Measured, for our own tests: on a 10-second and a 60-second consented clip, listeners in a blind test judge the clone the same speaker at least as often as the vendor's web-app clone. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: clone my voice from this recording
, make a voice from this clip
, create a custom voice from my sample
, upload this audio and give me a voice id
Does the job, with a limit 12
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Cartesia Voice Cloning API + MCP C Cartesia AI, Inc. | POST /voices/clone takes a clip (10 to 60 seconds on Sonic 3.6, 16 MB, flac, mp3 and others), a name and a required language, and returns a voice id in one call. No preprocessing or named out-of-limit error is documented, and Free has no cloning. source |
| ElevenLabs Voice Cloning and Voice Design API A Eleven Labs Inc. | POST /v1/voices/add takes files, a name, labels and an optional remove_background_noise, and returns a voice id with requires_verification. The reference states no sample length, size or format limits and only a generic 422, and acc-24 found a low-plan instant clone missed the speaker's inflection. source |
| Fish Audio Voice Cloning API B Hanabi AI Inc. | POST /model takes voices, optional matching texts (ASR runs otherwise), enhance_audio_quality (default true) and train_mode fast for an instant model id. No sample length or named out-of-limit error was found, and acc-02 found API clones much worse than web-app clones from the same audio. source |
| Gradium Voice Cloning and Voice Design B Gradium SAS | POST /api/voices/ takes an audio file, name, required language (en, fr, de, es or pt), start_s and enhance (default true, switchable off), and returns a voice id. About 10 seconds is recommended, but no maximum length, size or named out-of-limit error is stated. source |
| Inworld Instant Voice Cloning B Theai, Inc. dba Inworld AI | voices:clone takes one base64 sample (WAV or MP3, 3 to 30 seconds, 12 MB in the Portal guide), a langCode, an optional transcription and removeBackgroundNoise, and returns the voice in one call. Audio past 30 seconds is trimmed silently, not refused, and keys need write access. source |
| MiniMax Voice Clone and Voice Design C Nanonoble Pte. Ltd. | Upload with purpose voice_clone, then POST /v1/voice_clone with file_id and your own voice_id; limits are stated (10 seconds to 5 minutes, 20 MB, mp3, m4a or wav), with need_noise_reduction and named errors such as 2038 and 1043. A clone not used for synthesis within 7 days is deleted. source |
| Resemble AI Voice Cloning API D Resemble AI, Inc. | Create a voice with name, dataset_url (one WAV of 10 seconds to 3 minutes) and language, then call Build; training takes under a minute and callback_uri reports finished. Needs a Business plan or higher, and no preprocessing or out-of-limit error is documented. source |
| Smallest.ai Instant Voice Clone C Smallest, Inc. | One multipart POST with a 5 to 15 second clip under 5 MB (mp3, wav, mp4, webm), language and model returns voiceId and sample clips, with named errors. The default model is lightning-v3.1, whose clones stop on 30 November 2026; acc-31 (the founder) says clips must be clean. source |
| Soniox Voice Cloning BB Soniox Inc. | POST /v1/voices takes one clip up to 2 minutes and 35 MB; longer clips fail with voice_audio_too_long and bad files with named errors, and the voice is pollable until ready, typically in seconds. What preprocessing runs, and whether it matches the Console, is not stated. source |
| Speechify API Voice Cloning A Speechify, Inc. | POST /v1/voices takes a 10 to 30 second sample under 5 MB in any supported language with a consent recording and challenge id, and the voice is ready at once, with errors such as payload_too_large and unsupported_media_type. Preprocessing is not described, and the Free plan cannot clone. source |
| Ultravox Voice Cloning E Theai, Inc. dba Inworld AI | POST /api/voices takes one mp3 or wav file (30 to 60 seconds recommended, 10 MB maximum), a unique name and a description, and returns the voice. No language field, preprocessing or out-of-limit error is documented, and only one cloned voice per account is allowed. source |
| VoiceStudio D VoiceStudio maintainers | The MCP clone_voice takes base64 audio or a ref_audio_path and returns a profile_id for generate_speech. No sample length, format limits or preprocessing are stated in the MCP guide, and acc-09 found pressing clone did nothing in the Android app (closed). source |
Couldn't check 1
The docs were closed to us, or silent. These carry no mark either way.
| Listing | Why |
|---|---|
| Murf Voice Cloning API C Murf Inc. | Murf's terms bar automated access to murf.ai, so its API docs were not read this run, as in the 5 November check; no first-hand account names Murf. |
Doesn't do this job 1
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| Hume Octave Voice Design and Cloning + MCP C Hume AI, Inc. | Hume's docs index now lists only Audio Decisions and Expression Measurement, and the voice overview and changelog redirect there; the TTS and EVI APIs, cloning and design included, are no longer documented after the 13 November 2026 end of access. source |
What stops agents finishing
- no cloning on the free plan (Cartesia, Speechify, Gradium commercial use)
- cloning enabled per workspace by the vendor (Murf 403 until enabled) or Enterprise only (Hume, Resemble Business)
- API applies different preprocessing from the web app, so the API clone is worse (acc-02)
- only clean, single-speaker audio works and the clip the user holds is noisy (acc-28, acc-31)
Around it
Before: file-storage.read-file, speech-to-text.long-recording-batch. After: text-to-speech.clone-voice-speak, text-to-speech.long-form-narration.
Where this comes from
Weight 26 of 100 in Voice cloning & custom voices: 11 of 34 accounts, the largest group: people cloning their own or a character's voice and hitting sample limits, noisy references, accent drift, a cheap-plan cloner that fell short, and an API clone worse than the same vendor's web app
The same as data: /api/v1/use-cases/voice-cloning.clip-to-voice-id.json, or the MCP tool get_use_case with id voice-cloning.clip-to-voice-id. Other jobs in this category: /api/v1/niches?category=voice-cloning.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.