Comparison and Analysis of New Curriculum Criteria for End-to-End ASR (Aalto)
https://www.isca-speech.org/archive/interspeech_2022/karakasidis22_interspeech.html
https://github.com/aalto-speech/speechbrain-cl
The recommendation is to train on low WER samples first, then on high WER. However, the possible speedup is just about 2 times (-41%), which kind of destroys the promise of active learning. No order of magnitude win here.