Datasets:
Clarification on the Dataset Split Used for the Cleaned Version
It is claimed that your cleaned version is based on the test split of esb/datasets. I downloaded the esb/datasets data and found that the six provided audios are actually part of the training set. Could you please confirm whether this is correct?
I can confirm this independently. The esb/datasets loader (current revision 43a4ac0c, datasets.py) has no train list. It reads every call ID from
metadata.csv
(125 calls) and sets train = all_ids - _EARNINGS_DEV_IDS - _EARNINGS_TEST_IDS:
_EARNINGS_DEV_IDS = {4420696, 4448760, 4461799, 4469836, 4473238, 4482110}
_EARNINGS_TEST_IDS = {4432298, 4450488, 4470290, 4479741, 4483338, 4485244}
All 6 calls in this dataset (4468919, 4471586, 4474955, 4475604, 4482968, 4483046) are in the resulting ESB train split. None are in ESB dev or test. The ESB earnings22 test config in hf-audio/open-asr-leaderboard contains exactly the 6 test IDs above. The Open ASR Leaderboard uses the chunked version of this set (Earnings22-Cleaned-AA-chunked) in its default Average WER. Any model trained on ESB Earnings-22 train has seen this audio, and its score on this column will be optimistic.
Could you either correct the card, which currently says "Subset of Earnings-22 (ESB) English test split", or consider a version based on the ESB test calls listed above?