lochenchou/MOSNet
lochenchou/MOSNet has 381 GitHub stars and 63 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-19 19:14 UTC. Last push 2024-07-21.
Crecimiento de estrellas medido
- in 24h
- 0
- in 7 days
- Recopilando historial
- in 15 days
- Recopilando historial
- in 1 month
- Recopilando historial
Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.
Acerca de
Implementation of "MOSNet: Deep Learning based Objective Assessment for Voice Conversion"
Aprendizaje automáticoREADME.md
OriginalMOSNet
Implementation of "MOSNet: Deep Learning based Objective Assessment for Voice Conversion" https://arxiv.org/abs/1904.08352
Dependency
Linux Ubuntu 16.04
- GPU: GeForce RTX 2080 Ti
- Driver version: 418.67
- CUDA version: 10.1
Python 3.5
- tensorflow-gpu==2.0.0-beta1 (cudnn=7.6.0)
- scipy
- pandas
- matplotlib
- librosa
Environment set-up
For example,
conda create -n mosnet python=3.5
conda activate mosnet
pip install -r requirements.txt
conda install cudnn=7.6.0
Usage
Reproducing results in the paper
cd ./dataand runbash download.shto download the VCC2018 evaluation results and submitted speech. (downsample the submitted speech might take some times)- Run
python mos_results_preprocess.pyto prepare the evaluation results. (Runpython bootsrap_estimation.pyto do the bootstrap experiment for intrinsic MOS calculation) - Run
python utils.pyto extract .wav to .h5 - Run
python train.py --model CNN-BLSTMto train a CNN-BLSTM version of MOSNet. ('CNN', 'BLSTM' or 'CNN-BLSTM' are supported in model.py, as described in paper) - Run
python test.pyto test on the pre-trained weights with specified model and weight.
Note
The experimental results showed in the paper were trained on Keras with tensorflow 1.4.1 backend. However, the implementation here is based on tf2.0.0b1, so the results might vary a little. Additionally, the architectures showed in the paper were meta-architectures, any replace CNN/BLSTM with more fancy modules (ResNet etc.) would improve the final results. Tuning the hyper-parameters might result in the same favour.
Evaluating your custom waveform samples
- Put the waveforms you wish to evaluate in a folder. For example,
<path>/<to>/<samples> - Run
python python ./custom_test.py --rootdir <path>/<to>/<samples>
This script will evaluate all the .wav files in <path>/<to>/<samples>, and write the results to <path>/<to>/<samples>/MOSnet_result_raw.txt. By default, the pre_trained/cnn_blstm.h5 pretrained model is used. If you wish to use other models, please specify a different --pretrained_model and also change from model import <model_to_be_used>.
Citation
If you find this work useful in your research, please consider citing:
@inproceedings{mosnet,
author={Lo, Chen-Chou and Fu, Szu-Wei and Huang, Wen-Chin and Wang, Xin and Yamagishi, Junichi and Tsao, Yu and Wang, Hsin-Min},
title={MOSNet: Deep Learning based Objective Assessment for Voice Conversion},
year=2019,
booktitle={Proc. Interspeech 2019},
}
License
This work is released under MIT License (see LICENSE file for details).
VCC2018 Database & Results
The model is trained on the large listening evaluation results released by the Voice Conversion Challenge 2018.
The listening test results can be downloaded from here
The databases and results (submitted speech) can be downloaded from here