This is what I explored and worked with new SOMOS dataset during my Independent Project at MIDAS Lab

Trained MOS prediction models on different datasets using self-supervised learning model for speech, wav2vec2. Achieved a 45% improvement in MOS prediction accuracy compared to SOTA models on the SOMOS dataset.

Also leveraged Tacotron2 with TOBI labels to identify prosodic variations which give better naturalness scores (LCC,SRCC)

Data and pretrained models on Zenodo

Implentation of the PRS paper.

Download the data and pre-trained weights from zenodo in a folder named ASRU. We use the full path to ASRU folder to set up the REF variable later. We only provide PRS pre-trained weights for stage2.

Create a conda environment using the env.yaml conda env create -f env.yaml

Set the environment variable in linux based machine using this export REF="path/to/the/ASRU/"