SabriUgan_V2 RVC Model
RVC v2 tabanlı voice conversion modeli. Bu repo, farklı eğitim aşamalarındaki checkpointleri karşılaştırmak ve en doğal/kararlı sonucu seçmek için hazırlanmıştır.
Included Files
SabriUgan_V2_175e_10850s.pth
SabriUgan_V2_200e_12400s.pth
SabriUgan_V2_225e_13950s.pth
SabriUgan_V2.index
Training Summary
Architecture: RVC v2
Sampling rate: 48kHz
Vocoder: HiFi-GAN
Pitch extraction: RMVPE
Speaker embedder: contentvec
Training precision: FP16
Optimizer: AdamW
Loss: Single-Scale Mel Loss
Training data: ~71 minutes clean voice data
Final comparison checkpoints: 175 / 200 / 225 epoch
Checkpoint Notes
175 epoch: strong naturalness candidate
200 epoch: balanced candidate
225 epoch: final high-similarity candidate
The best checkpoint should be selected by comparing all three with the same source audio and identical inference settings.
Usage
Use with Applio or a compatible RVC inference interface.
Recommended starting inference settings:
Pitch extraction: RMVPE
Index file: SabriUgan_V2.index
Index rate: 0.75 - 0.90
Protect: ~0.33
Filter radius: 3
Resample: 0 or 48000
For best results, use a clean source recording with strong performance, natural pacing, and the intended emotion/intonation. RVC preserves the source delivery and converts the voice timbre.
Intended Use
This model is intended for internal testing, research, and controlled voice conversion workflows. Do not use it for impersonation, deception, public misuse, or any usage that violates rights, consent, or platform rules.
Status
Current comparison set is ready:
175e / 200e / 225e + index
Next step: run identical inference tests across the three checkpoints and select the final model.