RecursiveSynthVC

An expressive voice conversion model that is able to perform cross-speaker style transfer improved by self-generated synthetic expressive data.

TODO's

• Melspectrogram-based for lightweight training and explicit duration control
• BigVGAN V2 generator
• Large Scale Training for zero-shot voice conversion

Credits/Code resources and references

• VITS2 (https://github.com/p0p4k/vits2_pytorch/)
• NVIDIA BigVGAN (https://github.com/NVIDIA/BigVGAN)
• Speaker Normalized Affine Coupling layer (SNAC) (https://github.com/hcy71o/SNAC)
• Features preparation and Cosine Similarity based Speaker GRL (https://github.com/PlayVoice/whisper-vits-svc)
• F0 estimation Torch CREPE (https://github.com/maxrmorrison/torchcrepe)