Selection of Layers from Self-supervised Learning Models for Predicting Mean-Opinion-Score of Speech

This is the official implementation for the results reported in the paper "Selection of Layers from Self-supervised Learning Models for Predicting Mean-Opinion-Score of Speech" accepted at IEEE ASRU 2025.

Authors: Xinyu Liang, Fredrik Cumlin Emails: hopeliang@icloud.com, fcumlin@gmail.com

Training

The framework is Gin configurable; specifying model and dataset is done with a Gin config. See examples in configs/*.gin. The SSL features should be pre-generated using generate_audio_features.py and the path to the extracted features should passed as an argument to the training script.

SSL fine-tune on MOS prediction

The code can be found in Speech_SSL_finetune_MOS.

Model checkpoints

All projection head model checkpoints can be found in Huggingface, with 5 runs for each layer of each SSL model on each of the three datasets.

Inference

1. Download the model checkpoints from the Google Drive link.
2. Run python inference.py --audio_path=<YOUR_AUDIO_PATH> --ssl_model=<SELECTED_SSL_MODEL> --ssl_layer=<SELECTED_SSL_LAYER> --ckpt=<SELECTED_MODEL_CKPT_FROM>.