ComParE22

This repository provides the code for running the official baselines and Aalto's systems for ComParE2022.

Tamás Grósz, Dejan Porjazovski, Yaroslav Getman, Sudarsana Reddy Kadiri, and Mikko Kurimo: "Wav2vec2-based Paralinguistic Systems to Recognise Vocalised Emotions and Stuttering" in Proceedings of the 30th International Conference on Multimedia, (Lisbon, Portugal), ACM, 2022.

@inproceedings{Grosz22ComParE,
author = {Tam\'as Gr\'osz and Dejan Porjazovski and Yaroslav Getman and Sudarsana Reddy Kadiri and Mikko Kurimo},
title = {Wav2vec2-based Paralinguistic Systems to Recognise Vocalised Emotions and Stuttering},
booktitle = {{Proceedings of the 30th International Conference on Multimedia}},
year = {2022},
address = {Lisbon, Portugal},
publisher = {ACM},
month = {October},
note = {to appear},
}

Getting the code

You can find the baseline code and instructions for each sub-challenge on a corresponding branch of this repository.

To run our wav2vec2-based systems, use the ft_w2v2_classification_compare2022.py script.

The VGGish + TCN system can be reproduced using the codes in the VGGish+TCN branch. The augmentation script (utils/data_augmentation.py) we used in our experiments can also be found in that branch.

Links to the wav2vec2 models mentioned in the paper

• $wav2vec2_{M}$: https://huggingface.co/facebook/wav2vec2-large-xlsr-53
• $wav2vec2_{M}^{*}$: https://huggingface.co/voidful/wav2vec2-xlsr-multilingual-56
• $wav2vec2_{M}^{de}$: https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-german
• $wav2vec2_{L}^{de}$: https://huggingface.co/aware-ai/wav2vec2-xls-r-1b-5gram-german
• $wav2vec2_{M}^{er}$: https://huggingface.co/superb/wav2vec2-large-superb-er
• $wav2vec2_{S}^{de}$: https://huggingface.co/aware-ai/wav2vec2-base-german