[GH-ISSUE #2025] Why does my RVC model produces bad results? #1366

New issue

Open

opened 2026-10-20 10:08:30 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-20 10:08:30 +00:00

Owner

Originally created by @GUUser91 on GitHub (Apr 29, 2024).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/2025

I created a RVC model and I had it set to 50 epochs for training. It produces bad results. Here is the audio file I used for training, I used version v2 and I set the target sample rate to 48k.
https://files.catbox.moe/440g5r.wav
Here is the input file I used.
https://vocaroo.com/11ANq5W9xFd5
Here is the output file.
https://vocaroo.com/136PewYoZmL3

Edit: I discovered Seed-VC which worked well with 6 minutes of training data. It was also worked with 14 seconds of training data.

Originally created by @GUUser91 on GitHub (Apr 29, 2024). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/2025 I created a RVC model and I had it set to 50 epochs for training. It produces bad results. Here is the audio file I used for training, I used version v2 and I set the target sample rate to 48k. https://files.catbox.moe/440g5r.wav Here is the input file I used. https://vocaroo.com/11ANq5W9xFd5 Here is the output file. https://vocaroo.com/136PewYoZmL3 Edit: I discovered Seed-VC which worked well with 6 minutes of training data. It was also worked with 14 seconds of training data.

bsakaguchi added the

help wanted

label

2026-10-20 10:08:30 +00:00

bsakaguchi commented

2026-10-20 10:08:30 +00:00

Author

Owner

@HotChocut commented on GitHub (Apr 30, 2024):

Your training data is good but you need more of it, 40 seconds is not enough to train a good RVC model. Aim for 10 minutes if possible but the more the merrier, I've done models with upwards of 4 hours worth of training data so there really is no such thing as "too much data".
I personally would also train with more epochs (around 200) but most of the time I just go with the shotgun approach and do 1000 which is DEFINITELY overkill and I would not recommend unless you're crazy like I am.

@HotChocut commented on GitHub (Apr 30, 2024): Your training data is good but you need more of it, 40 seconds is not enough to train a good RVC model. Aim for 10 minutes if possible but the more the merrier, I've done models with upwards of 4 hours worth of training data so there really is no such thing as "too much data". I personally would also train with more epochs _(around 200)_ but most of the time I just go with the shotgun approach and do 1000 which is **DEFINITELY** overkill and I would not recommend unless you're crazy like I am.