This is code for our paper which has been accepted to ICASSP 2022:
"Generalization Ability of MOS Prediction Networks" Erica Cooper, Wen-Chin Huang, Tomoki Toda, Junichi Yamagishi https://arxiv.org/abs/2110.02635
Please cite this preprint if you use this code.
- Fairseq toolkit: https://github.com/pytorch/fairseq Make sure you can
import fairseqin Python. - torch, numpy, scipy, torchaudio
- I have exported my conda environment for this project to
environment.yml - You also need to download a pretrained wav2vec2 model checkpoint. These can be obtained here: https://github.com/pytorch/fairseq/tree/main/examples/wav2vec If you are using the
run_inference_for_challenge.pyscript, one will be downloaded for you automatically. Otherwise, please choosewav2vec_small.pt,w2v_large_lv_fsh_swbd_cv.pt, orxlsr_53_56k.pt. - You also need to have a MOS dataset. You can find the BVCC dataset of MOS ratings that was used for the VoiceMOS Challenge here: https://zenodo.org/record/6572573#.Yphw5y8RprQ
Run python run_inference.py --datadir /path/to/your/wavdir
The wavdir should contain .wav audio files for which you want to predict MOS.
Note: These .wav files should already be downsampled to 16kHz and sv56-normalized. Please do this by yourself first.
The output will be answers.txt which will contain the name of each audio file and its predicted MOS.
CAVEAT EMPTOR: This MOS predictor is experimental technology. It was trained on English-language audio samples from text-to-speech and voice conversion systems from 2008-2020, with their MOS ratings for naturalness. It may not work well on audio from other models, languages, domains, speakers, etc. That being said, we would be really happy to hear about any ways that you are using this pretrained model, and how well it worked for your use case -- please let us know if you have comments or if you publish your results!
Please see instructions in VoiceMOS_baseline_README.md.
This study is supported by JST CREST grants JP- MJCR18A6, JPMJCR20D3, and JPMJCR19A3, and by MEXT KAKENHI grants 21K11951 and 21K19808. Thanks to the organizers of the Blizzard Challenge and Voice Conversion Challenge, and to Zhenhua Ling, Zhihang Xie, and Zhizheng Wu for answering our questions about past challenges. Thanks also to the Fairseq team for making their code and models available.
BSD 3-Clause License
Copyright (c) 2021, Yamagishi Laboratory, National Institute of Informatics All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
-
Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
-
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
-
Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.