fcumlin/lamosnet

LaMOSNet

Author: Fredrik Cumlin Email: [email protected]

This is the official implementation of the model LaMOSNet from the paper "Latent-based Neural Net for Non-intrusive Speech Quality Assessment". LaMOSNet is a non-intrusive speech quality metric that given an input signal, predicts the overall quality.

Training

First, download the VCC2018 data, which can for example be done from here. Then run train.py, which depend on the following arguments:

• --num_epochs: The number of epochs during training.
• --log_valid: The number of epochs between logging results on validation data.
• --log_epoch: The number of epochs between logging training losses.
• --data_path: Path to the data folder.
• --id_table: Path to the id_table folder.
• --save_path: Path for the best model to be saved.

Acknowledgement

This repository inherits from the unofficial MBNet implementation and the LDNet implementation .