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.

Notebooks/files to reproduce the experiments:

Experiment E1 - Noisy waveform input

• wav2vec2 vs HuBERT on LibriSpeech - noise1_wv2_vs_hubert_revised.ipynb
• wav2vec2 vs DistilHuBERT on TIMIT - final_timit_expt.ipynb

Experiment E2A - Layer noise injection

• Additive noise - final_white_noise_layers.ipynb, needs dependency custom_wav2vec2.py
• Multiplicative noise - final_multiplicative_noise_layers.ipynb, needs dependency custom_mult_wav2vec2.py

Experiment E2B - Layer activation visualization

• visualization_expts.ipynb