metadata
license: openrail
Test RVC models on the DDLC character Monika, via various hyperparams and datasets.
monika-test-0 (~07/2023)
- Trained on augmented dataset of ~10 10 second clips
- Trained for ~100 epochs
- RVC1
- "Version 1" ("0" in the old numbering)
monika-test-2 (~07/2023)
- Trained on augmented dataset of ~10 10 second clips (augmented via tortoise tts)
- Trained for 100 epochs
- RVC1
monika-test-4 (~07/2023)
- Trained on smaller but better dataset of ~2 10 second clips (augmented via 11labs)
- Trained for 150 epochs
- RVC1
monika-test-7 (08/22/2023)
- Trained on augmented dataset of ~10+ 10 second clips (augmented via tortoise tts)
- Trained for 60 epochs (720 steps)
- Better quality than others
- "Version 2" ("1" in old numbering)
- RVC2
monika-test-8 (08/22/2023)
- Trained on smaller but better dataset of ~5 10 second clips (some augmented via 11labs)
- Trained for 60 epochs (660 steps)
- Even clearer quality but with slightly more artifacting than monika-test-7 (still better than pre 7th ones)
- "Version 2a" ("1a" in old numbering)
- RVC2
ct-m3 (~10/2023)
- Trained on preprocessed version of dataset of ~5 10 second clips
- Trained for ~100 epochs
- Test model
- RVC1
ct-m4 (~10/2023)
- Trained on preprocessed version of dataset of ~5 10 second clips
- Trained for ~200 epochs
- Test model
- RVC1
ct-m4a (~10/2023)
- Trained on preprocessed version of dataset of ~5 10 second clips
- Trained for ~200 epochs
- "Version 4" ("3" in old numbering)
- RVC2
fused2 (~02/2024)
- Merge between ct-m3 and another model ("Sayori"-based model, with ratio of 75% to 25%)
- Somewhat clearer quality
- Yet another test model
- RVC2
fused5 (~01/2025)
- 50-50 merge between ct-m5 (~06/2024) and fused2
- In addition to merging, experiment with fine-tuning on more synthetic data which doubled preprocessed dataset size
- Sliiightly better than fused2; seems not to "break" where fused2 does, otherwise they seem almost the same
- RVC2