hewliyang/nisqa-blizzard-challenge-mos

blizzard-challenge-mos

This dataset contains MOS (Naturalness) subjective metrics derived from listening tests done during the evaluation stages of the Blizzard Challenge from 2008 -> 2023. All data is sourced from https://www.cstr.ed.ac.uk/projects/blizzard/data.html.

The dataset is formatted for training NISQA models. In particular, at this time there are only two columns for each .csv file:

• filepath_deg: relative file path to the .wav file
• mos: averaged MOS per file over an X number of listeners**

Usage

Similarly, we cannot leverage datasets for data loading. Instead, make sure Git LFS is installed on your system:

sudo apt install git-lfs
git lfs install

Download the dataset using:

git clone https://huggingface.co/datasets/hewliyang/blizzard-challenge-mos

Untar the wavfiles

find . -name '*.tar.gz' -exec tar -xzf '{}' \;

# can remove the tarballs to save disk space
rm *.tar.gz

Summary

Taken from Mittag. G, Moller. S, 2021

To filter for subtasks ie: EH1, EH2 notice that the filenames have been make extremely verbose by design.
You can filter easily by checking for substrings ie: "EH1" in x

Caveats**

1. 2014,2015,2016 data only contains system level (ie: submission grouped by participating team).
2. 2015 data was no longer available for download at the link above, hence has been omitted here.

Todos

• handle 2014 data (??)
• add metadata (license per system, languages, hub/spoke task labels)
• handle 2016 edge cases (listening tests were done on a concatenation of a few raw files + silences. these samples have to be reconstructed which has not been done yet. so we are missing ~500 wavs)
• citations for all Blizzard Challenges

References

@inproceedings{Mittag_2021, series={interspeech_2021},
   title={NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets},
   url={http://dx.doi.org/10.21437/Interspeech.2021-299},
   DOI={10.21437/interspeech.2021-299},
   booktitle={Interspeech 2021},
   publisher={ISCA},
   author={Mittag, Gabriel and Naderi, Babak and Chehadi, Assmaa and Möller, Sebastian},
   year={2021},
   month=aug, collection={interspeech_2021} }

@inproceedings{Mittag_2020, series={interspeech_2020},
   title={Deep Learning Based Assessment of Synthetic Speech Naturalness},
   url={http://dx.doi.org/10.21437/Interspeech.2020-2382},
   DOI={10.21437/interspeech.2020-2382},
   booktitle={Interspeech 2020},
   publisher={ISCA},
   author={Mittag, Gabriel and Möller, Sebastian},
   year={2020},
   month=oct, collection={interspeech_2020} }