Skip to content

Official implementation of the paper, "Opening the Black Box of wav2vec Feature Encoder."

License

Notifications You must be signed in to change notification settings

juice500ml/unbox-w2v-convnet

Repository files navigation

Opening the Black Box of wav2vec Feature Encoder

Official implementation of the paper: https://arxiv.org/abs/2210.15386

How to run

# This will generate *.pkls to the pkls/ folder, saving convolutional features.
python3 save_embeddings.py

# Then, use jupyter to open visualization.ipynb
# It should generate plots without any problem!
jupyter lab

# FYI, you may try different hyperparameter grids via fixing configs.py

Praat experiments

You need a Praat installation beforehand to run the praat/generate_audio.praat via automated script, praat/grid_audio.sh. You'll have to uncomment some of the code in configs.py before saving the embeddings.

Acknowledgements

About

Official implementation of the paper, "Opening the Black Box of wav2vec Feature Encoder."

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages