I Used Replay to Train My Own Voice Model!
Вставка
- Опубліковано 28 вер 2024
- Replay is software (and it's free!) that allows you to replace vocals using voice models. It's one of the ways people use to make AI covers. It's fun to play around with, and I've done that before using Hololive voice models (just be sure you are clearly stating that it's not the real thing).
But I've always wanted to train my own voice and create my own voice model! The problem was, when I looked into it, it all seemed so complicated and I'm just very stupid. However, Replay has added (beta) feature to Train Your Own Model! Cool! So I did that, training it on about 10 minutes worth of me narrating something -- this was a recording I had done a few years ago.
I tested it by swapping my voice in using an example where Microsoft AI TTS read some news. It really does sound like me! Next, I wanted to see how I sounded singing. So I created clips using Synthesizer V Xuan Yu and Jun as the base vocals and I compared that to my actual singing voice.
I could tell that it was my voice, but the thing is, my talking voice is not the same as my singing voice. What I want to do next is train a voice model using clips of me just singing. I'll have to generate isolated stems from early recording projects for that.
Exactly what I have been trying to achieve. Right on, thank you.
1:01 hey you're actually pretty smart have some confidence in yourself. I know next to nothing about a lot of this stuff other than it's really cool so you've got me beat for sure.
What model of Gpu do you have? Or did you use cpu? I have a 3050 and with a 2 minute sample it is taking forever!
I have the Geforce RTX 3070. Training 12 minutes took 3 hours and 45 minutes.
@@AnomieTrain Thanks! No wonder it was taking so long!
I'm using a 2060 and it's still going ... just clicked over 5 hours !!!
You've got me thinking. I've been using Suno then separating the vocal, putting it through musicfy using my trained vocal and it was terrible but also terrible with their trained models. So, obviously it was because of the poor quality vocal separation. Now I'm going to import the separated vocal audio into Synth v and use the output from that to change it to my voice. Going to use replay because I cancelled my musicfy sub. Thanks.