Audio-WestlakeU/VINP

Combines variational Bayesian inference with a neural speech prior DNN to jointly perform speech dereverberation and blind room impulse response (RIR) identification on reverberant audio. The method leverages a probabilistic graphical model based on cepstral-temporal filtering (CTF) to extract the anechoic speech prior and RIR estimates without requiring joint training with downstream ASR systems. Supports multi-GPU distributed training via PyTorch's torchrun, with evaluation tools for speech quality metrics, ASR performance (via Whisper), and RIR acoustic parameters (RT60/DRR).