codename0og/codename-rvc-fork-4

Supports multiple custom vocoder architectures (RefineGAN, RingFormer, PCPH-GAN) alongside standard HiFi-GAN, with configurable training via hold-out validation metrics (PESQ, SI-SDR, mrSTFT), multiple optimizers (AdamW, RAdam, Prodigy), and adversarial loss variants. Includes ML-based silence truncation, VITS2 enhancements with speaker-conditioned text encoding, and per-epoch loss averaging with rolling averages for improved training transparency. Provides Gradio UI with quick parameter tweaks (learning rates, schedulers, KL annealing) and TensorBoard integration for real-time monitoring.