How to train your own AI voice model

What you need

• 3 to 60 minutes of one voice. 3 minutes is the minimum Jammable accepts; voices usually sound better with 6 minutes or more. Only singing or speech counts: silence is cut automatically and a file added twice only counts once.
• Clean recordings. One voice at a time, no backing vocals, no crowd, no music behind it, in a quiet room without echo.
• Range. Different notes, words and energy levels. A voice trained only on quiet speech will struggle to sing loudly.
• MP3, WAV or M4A files, or just your microphone: you can record straight into the page.

Use recordings of your own voice, or of someone who has agreed to it.

Step by step

1. 1

   Open Train and add your voice

   Go to Train and upload files, or press Record and sing or talk. Record as many takes as you like.

2. 2

   Check what gets kept

   Each file shows how much voice is left after the silence is removed. Trim anything that isn't the voice you want, like a cough, a second person or a noisy intro.

3. 3

   Name it and choose who can use it

   Add a name and a picture, then pick Private, only me or Public in the library.

4. 4

   Clean up the samples if needed

   Turn on Remove instrumentals if there is music under the vocals, and Remove reverb if the room echoes. Skip both for dry, clean recordings.

5. 5

   Start training

   It runs on Jammable's servers, so you can close the page. You get an email when the voice is ready.

6. 6

   Use it

   Your voice appears in your Library. Pick it on the cover maker or for text-to-speech like any other voice.

How long training takes

Tips for a better voice

• Quality beats quantity. 8 clean minutes train a better voice than 40 minutes recorded in a noisy room.
• Record the way you will use it. If the voice is for singing, give it singing, across your low and high notes. For text-to-speech, read varied sentences out loud.
• Keep the microphone in one place. The same distance and the same room for every take keep the tone consistent.
• Don't stack effects. Autotune, heavy compression and reverb get learned as part of the voice.
• Retrain rather than patch. If the result sounds thin or robotic, add more varied audio and train again instead of tweaking the cover settings.

Training on your phone or computer

Training happens in the browser on any device, so there is nothing to install and no graphics card needed. People who run RVC on their own computer can also upload an existing RVC model (a .pth file plus its .index file) instead of training a new one.