This is the repository containing the code for the paper Analyzing robustness of end-to-end neural models for automatic speech recognition.
Access it at https://arxiv.org/abs/2208.08509
Slides for our work is available at presentation/presentation.pptx
If you have comments or suggestions, please reach out to weizou@uchicago.edu or goutham@uchicago.edu.
- wav2vec2 vs HuBERT on LibriSpeech -
noise1_wv2_vs_hubert_revised.ipynb
- wav2vec2 vs DistilHuBERT on TIMIT -
final_timit_expt.ipynb
- Additive noise -
final_white_noise_layers.ipynb
, needs dependencycustom_wav2vec2.py
- Multiplicative noise -
final_multiplicative_noise_layers.ipynb
, needs dependencycustom_mult_wav2vec2.py
visualization_expts.ipynb