How to train your own AI voice model

By the voicechange.io team · November 6, 2026

You can make an AI model of your own voice for free with the VoiceChange desktop app. Then you can hear text in your voice, or make song covers that sound like you singing. The quality of the model depends almost entirely on the recordings you give it.

What "training a voice" means

Training builds a new RVC model that has learned the timbre of your voice: the tone, texture and resonance that make you recognisable. It does not learn your words or your accent; those always come from whatever audio or text you convert later. Think of it as teaching the system what your voice sounds like, not what it says.

What you need

  • 2 to 10 minutes of audio in total. You can upload several files.
  • One speaker only: just you, no interviewer, no one else in the room.
  • No music or background noise. This matters more than anything else.
  • Files up to 100 MB in total, in any common audio format.
  • The right to use the voice. Only train on your own voice, or a voice whose owner has given you permission.

Recording a good dataset

The model will reproduce whatever it hears, including room echo, hum and mouth clicks. Spend ten minutes getting the recording right and you'll save yourself a disappointing model.

  1. Pick a quiet, soft room. A small room with soft furnishings (a bedroom, or even a wardrobe full of clothes) is ideal. Turn off fans and close the windows.
  2. Use the same mic for everything, about a hand's width from your mouth. A recent phone is perfectly good.
  3. Cover your range. Read a few paragraphs normally, then some lines with more energy, some quieter, a few questions and exclamations. If you want to use the model for song covers, include a couple of minutes of singing too, ideally a cappella.
  4. Be natural. Speak the way you want the model to sound. A monotone read gives a monotone voice.
  5. Trim the junk. Cut out coughs, long silences, laughter from others and any moments where something loud happened in the background.
More is not always better. Five minutes of clean audio beats twenty minutes of noisy audio. If you only have a few great minutes, use only those.

Where to clone your voice

Voice cloning is in the free VoiceChange desktop app for Windows, macOS and Linux. It is no longer done on the website. Use the advice above to prepare your recordings before you start.

If you trained a voice on the website before, it is still in the Your voice tab of the voice changer, in the same browser you used. You can keep using it there or delete it. Your original recordings were deleted when training finished.

If the result isn't good enough

  • Sounds noisy or metallic: the training audio had background noise. Re-record in a quieter room.
  • Doesn't sound like you: usually too little audio, or audio from several different mics. Record 5+ minutes in one session.
  • Fine for talking, poor for singing: the model has never heard you sing. Add singing to the dataset and train again.
  • Training failed: the most common cause is files that are mostly silence or music. Check your files and try again.

Once you have a model, the recording tips apply when converting with it too.

Try it now: the voice changer is free and needs no account. Pick from all voices, or read the FAQ.

More guides