Hear what a vocal split leaves you with
This is a real 4-stem result from the engine behind the tool. Play the Vocals player on its own. That is roughly the raw material you get from a song: the voice without drums, bass or synths, but still with the reverb that was on the voice in the mix. Judge your own file the same way before you feed it to a cloner.
A spoken recording with background music, such as a vlog or a podcast intro, usually gives a cleaner voice sample than a sung one. Speech has fewer held notes and less reverb, and most cloning tools are built for speech.
What popular cloning tools ask for
Each tool publishes its own sample rules. These are taken from their documentation, checked on 31 October 2026. Follow the tool you will actually use, because the numbers differ a lot.
| Tool | Audio length | Sample rules | Format | Consent |
|---|---|---|---|---|
| ElevenLabs Instant Voice Clone | 1 to 2 minutes recommended, at least 10 seconds | One speaker, no background noise or reverb, about -23 to -18 dB RMS, peaks under -3 dB | MP3 at 128 kbps or higher | You confirm you have the rights and consent before saving |
| ElevenLabs Professional Voice Clone | At least 30 minutes, 2 to 3 hours recommended | Speech only: singing and music are excluded | MP3 at 192 kbps or higher | Your own voice only, verified by reading lines aloud |
| Fish Audio | At least 10 seconds per clip, 1 to 2 minutes for better fidelity | Clean, mono, single speaker, no music, reverb or overlapping voices | WAV, MP3, M4A or Opus | Not covered in the model docs |
| RVC (open source) | 10 to 50 minutes recommended, 5 to 10 minutes can work if clean | Low noise, consistent voice; under 1 minute not reported to work | WAV preferred by community guides | Up to you and the law where you live |
Sources: ElevenLabs voice cloning docs, Fish Audio developer guide, and the RVC project FAQ on GitHub. ElevenLabs also says its own Voice Isolator is not optimised for vocals from music, which is the case a music separation model like Demucs is trained for.
Prep checklist, step by step
- Isolate the voice. Use the 2-stem split above to remove drums, bass and other instruments.
- Deal with reverb separately. The split keeps the reverb that is on the voice. If your tool wants dry audio, run a de-reverb plugin in your editor, or pick drier passages.
- Denoise lightly. Heavy noise reduction makes a voice sound robotic, and the clone copies that.
- Keep one speaker and no harmonies. Cut backing vocals, duets, ad-libs and crowd parts.
- Remove long silences, but keep natural pauses between sentences.
- Level it. ElevenLabs suggests about -23 to -18 dB RMS with peaks under -3 dB; normalise in a free editor such as Audacity.
- Export to the format your tool asks for. Mono is fine for most; Fish Audio asks for mono.
Community RVC guides follow the same order: separate the vocal, remove echo and reverb, denoise, then slice into clips of about 10 seconds. Each processing pass can cost a little quality, so do only the steps your recording needs.
Singing vs speaking samples
| Sample source | Works for | Watch out for |
|---|---|---|
| Vlog, podcast or voice-over with music under it | Speech clones (ElevenLabs, Fish Audio) | Music bleed in quiet gaps, room noise |
| A song you sang | Singing voice models such as RVC | Harmonies, reverb, held notes; ElevenLabs PVC rejects singing |
| Clean studio voice memo | Any tool, no split needed | Nothing to separate: upload it as is |
If you can record fresh audio, do that. A quiet room and a plain reading of text beats any cleaned-up song. The split is for when your best or only recording has music in it.
Consent comes first
Only clone your own voice, or a voice whose owner has given you clear, written permission. ElevenLabs asks you to confirm rights and consent for every instant clone and limits professional clones to your own voice. Cloning a singer or a public figure from their records without permission can break the tool's terms and the law, and it is not what this page is for.
TuneSplit stores uploads privately, uses them only to make your stems, deletes them within 24 hours and never uses them to train a model, including any voice model.
More tools for the same job
- Remove music from video: keep the voice from an MP4 or MOV.
- Acapella extractor: vocals-only files for remixes and DJ edits.
- Vocal remover: the reverse job, music without the voice.
- Audio cutter: trim silences and bad takes in your browser.
Frequently asked questions
How much audio do I need for AI voice cloning?
It depends on the tool. ElevenLabs Instant Voice Cloning recommends 1 to 2 minutes and accepts as little as 10 seconds. Its Professional clone needs at least 30 minutes. Fish Audio works from 10 seconds, and the RVC project recommends 10 to 50 minutes.
Can I clone a voice from a song?
You can prepare a sample from a song you sang yourself: split off the music, cut out harmonies and keep the clean lines. Singing voice models such as RVC can use it. ElevenLabs Professional Voice Cloning excludes singing, so record speech for that.
Does TuneSplit clone voices?
No. TuneSplit separates audio into stems, so it gives you the voice without the music. You then take that file to a voice cloning tool of your choice. Uploads are deleted within 24 hours and never used to train any model.
Does the vocal split remove reverb?
No. The split removes instruments, but reverb on the voice stays with the voice. If your cloning tool asks for dry audio, use a de-reverb plugin in your editor or choose the driest passages of the recording.
What format should a voice cloning sample be?
Follow your tool. ElevenLabs asks for MP3 at 128 kbps or higher for instant clones and 192 kbps for professional ones. Fish Audio accepts WAV, MP3, M4A and Opus. Community RVC guides prefer WAV. The free vocal split here downloads as MP3.
Is it legal to clone someone else's voice?
Only with their permission. Cloning tools ask you to confirm you have the rights and consent, and some only allow your own voice. Using another person's voice without consent can break the tool's terms and laws on likeness and fraud where you live.
Why does my clone sound robotic or have music in it?
Usually the sample is the problem: leftover music, heavy noise reduction, reverb or two voices at once. Re-cut the sample to clean single-voice passages, denoise lightly and level it to about -23 to -18 dB RMS before you try again.