Whispering Wisely: Smart Subset Selection for ASR Fine-Tuning

Researchers propose an embedding-based subset selection method for fine-tuning ASR models like OpenAI's Whisper, using a greedy k-center algorithm on Whisper's pretrained encoder embeddings to select diverse training subsets. The approach is evaluated on Hebrew and Danish, showing k-center selection often outperforms random sampling, especially at smaller subset sizes, and captures broader lexical diversity.