SQUIM Report

This is an usage implementation of TorchAudio-SQUIM. The end goal is to create a dataset report generator based on the speech quality objective metrics - STOI, PESQ, SI-SDR. May consider adding other metrics in the future.

TODOs

• Script to estimate SQUIM for a single audio file
• Gradio demo for estimating SQUIM a single audio file
• Dataset report generator

Install

Note: TorchAudio-SQUIM is still in development, and only available in TorchAudio's main branch. You will have to install the preview (nightly) build to use it. Run the command below to install the nightly build for CUDA 11.8 using pip, or head over to pytorch.org for other versions.

# CUDA 11.8 nightly build
pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
# for Gradio demo
pip3 install gradio
# Clone repository
git clone https://github.com/test-dan-run/squim-report.git

How to Use

For a single file.

python3 infer.py <path-to-audio-file>
# [test.wav] SQUIM scores: {'STOI': 0.677, 'PESQ': 1.194, 'SI-SDR': -1.396}

Citations

@INPROCEEDINGS{10096680,
  author={Kumar, Anurag and Tan, Ke and Ni, Zhaoheng and Manocha, Pranay and Zhang, Xiaohui and Henderson, Ethan and Xu, Buye},
  booktitle={ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Torchaudio-Squim: Reference-Less Speech Quality and Intelligibility Measures in Torchaudio}, 
  year={2023},
  volume={},
  number={},
  pages={1-5},
  doi={10.1109/ICASSP49357.2023.10096680}}