MOS_Estimation2

Apply a pre-trained model to wav files to obtain a MOS quality score Based on work reported in: "Comparison of Speech Representations for Automatic Quality Estimation in Multi-Speaker Text-to-Speech Synthesis" https://arxiv.org/abs/2002.12645 ( Note this code was derived from https://github.com/lochenchou/MOSNet, 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

1. place all speech wavfiles into a directory, AUDIO_DIR
2. Run python utils.py AUDIO_DIR to prepare the wav files.
3. Run python apply_model.py to generate MOS scores
4. Final results will be provided in a CSV file, in data/results.txt, as well as a pickle format (useful for plotting, using origial MOSnet code) data/results.pkl

Note a small example using VCTK data is provided, the wav files are in the directory audio/. Also note that this example is set up to parse speaker ID from the filename. You can change this, based on your own file name convention, as well as using different systems, by modifying line 76 in apply_model.py