Author: Wen-Chin Huang (Nagoya University) Email: [email protected]

This is the official implementation of the paper "LDNet: Unified Listener Dependent Modeling in MOS Prediction for Synthetic Speech". This is a model that takes an input synthetic speech sample and outputs the simulated human rating.

• PyTorch 1.9 (versions not too old should be fine.)
• librosa
• pandas
• h5py
• scipy
• matplotlib
• tqdm

The following instructions are for the VCC2018 benchmark.

New: This system is also one of the baseline systems of the first VoiceMOS Challenge. Please refer to this document for detailed instructions.

# Download the VCC2018 dataset.
cd data
./download.sh vcc2018

We provide configs that correspond to the following rows in the above figure:

• (a): MBNet.yaml
• (d): LDNet_MobileNetV3_RNN_5e-3.yaml
• (e): LDNet_MobileNetV3_FFN_1e-3.yaml
• (f): LDNet-MN_MobileNetV3_RNN_FFN_1e-3_lamb4.yaml
• (g): LDNet-ML_MobileNetV3_FFN_1e-3.yaml

python train.py --config configs/<config_name> --tag <tag_name>

By default, the experimental results will be stored in exp/<tag_name>, including:

• model-<steps>.pt: model checkpoints.
• config.yml: the config file.
• idtable.pkl: the dictionary that maps listener to ID.
• training_<inference_mode>: the validation results generated along the training. This file is useful for model selection. Note that the inference_mode in the config file decides what mode is used during validation in the training.

There are some arguments that can be changed:

• --exp_dir: The directory for storing the experimental results.
• --data_dir: The data directory. Default is data/vcc2018.
• seed: random seed.
• update_freq: This is very important. See below.

By default, all LDNet models are trained with a batch size of 60. In my experiments, I used a single NVIDIA GeForce RTX 3090 with 24GB mdemory for training. I cannot fit the whole model in the GPU, so I accumulate gradients for update_freq forward passes and do one backward update. Before training, please check the train_batch_size in the config file, and set update_freq properly. For instance, in configs/LDNet_MobileNetV3_FFN_1e-3.yaml the train_batch_size is 20, so update_freq should be set to 3.

python inference.py --tag LDNet-ML_MobileNetV3_FFN_1e-3 --mode mean_listener

Use mode to specify which inference mode to use. Choices are: mean_net, all_listeners and mean_listener. By default, all checkpoints in the exp directory will be evaluated.

There are some arguments that can be changed:

• ep: if you want to evaluate one model checkpoint, say, model-10000.pt, then simply pass --ep 10000.
• start_ep: if you want to evaluate model checkpoints after a certain steps, say, 10000 steps later, then simply pass --start_ep 10000.

There are some files you can inspect after the evaluation:

• <dataset_name>_<inference_mode>.csv: the validation and test set results.
• <dataset_name>_<inference_mode>_<test/valid>/: figures that visualize the prediction distributions, including;
  • <ep>_distribution.png: distribution over the score range (1-5).
    
  • <ep>_utt_scatter_plot_utt: utterance-wise scatter plot of the ground truth and the predicted scores.
    
  • <ep>_sys_scatter_plot_utt: system-wise scatter plot of the ground truth and the predicted scores.
    

This repository inherits from this great unofficial MBNet implementation.

If you find this recipe useful, please consider citing following paper:

@article{huang2021ldnet,
  title={LDNet: Unified Listener Dependent Modeling in MOS Prediction for Synthetic Speech},
  author={Huang, Wen-Chin and Cooper, Erica and Yamagishi, Junichi and Toda, Tomoki},
  journal={arXiv preprint arXiv:2110.09103},
  year={2021}
}