Agha Ali Raza

Publications/Efficient Machine Learning: Data Selection, Pruning, Active & Federated Learning

Workshop · 2023

Representative Subset Selection for Efficient Fine-Tuning in Self-Supervised Speech Recognition

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza

NeurIPS 2023 Workshop (ENLSP), New Orleans, USA

Summary

Fine-tuning self-supervised speech recognition models requires large amounts of labeled audio. Which subset should be used when resources are limited? Surprisingly, pruning methods designed for images do no better than random selection here. We introduce COWERAGE, which picks a representative subset that covers the full range of difficulty, measured by each example's word error rate early in training. With wav2vec 2.0 and HuBERT on three datasets, it improves word error rate by up to 17% relative to existing pruning methods and random sampling, because covering that range also brings in phonemically diverse examples.

Cite this paper

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza. Representative Subset Selection for Efficient Fine-Tuning in Self-Supervised Speech Recognition. NeurIPS 2023 Workshop (ENLSP), New Orleans, USA.

@inproceedings{azeemi2023representative,
  author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{Representative Subset Selection for Efficient Fine-Tuning in Self-Supervised Speech Recognition}},
  booktitle = {NeurIPS 2023 Workshop (ENLSP), New Orleans, USA},
  year = {2023},
  url = {https://aghaaliraza.com/publications/2023-NeurIPSWS--Representative_Subset_Selection_for_Efficient_Fine-Tuning_in_Self-Supervised_Speech_Recognition.pdf}
}

More on this theme

All 8 works on this theme →