Demo for MusDB vocals

Audio Reconstruction from Mel-spectrogram using different models tested on MusDB vocals.

• Vocos: Model Trained on Libritts
• PriorGrad: Model Trained on Libritts
• FreGrad: Model Trained on Libritts
• BigVGAN-base: Model Trained on Libritts, 14M Parameters
• BigVGAN: Model Trained on Libritts, 112M Parameters
• RFWave: Model Trained on Libritts, enabling STFT Loss
• BigVGAN-V2: Model Trained on Nvidia's Large-scale Compilation Dataset, 112M Parameters
• RFWave-Universal: Model Trained on DNS-Chanllenge, MTG-Jamendo and AudiuoSet, enabling STFT Loss

Spectrogram example pictures