Datasets:
Label mismatch
First, I truly appreciate you sharing this public dataset, which will help push the ASR field forward.
However, I noticed what seems to be a conflict between the human and ASR transcripts in the "Mongolian (mn-mn)" subset. When I checked the actual audio against both transcripts, the ASR transcript was actually more accurate, which is both impressive and surprising.
From my perspective, there might be a column-parsing error during the dataset processing. The paper mentions using the bayartsogt/whisper-large-v2-mn-13 model to transcribe the Mongolian audio, but it usually isn't that accurate compared to ground-truth human data.
Could the authors please double-check the two labeled columns and clarify which one is actually the human transcription?
Thank you again for all your hard work!
Hi @onlysainaa , thank you for the careful check and the kind words! We double-checked the mn_mn columns and they are not swapped: human_transcript really is the official parliamentary record, and asr_transcript is the Whisper output. What you observed comes from two properties of how the dataset is built:
- Segment boundaries. The human transcript for each segment is recovered by sliding-window character-level alignment against the official record, so the first/last words of a segment can occasionally be clipped or include a small fragment of the neighbouring sentence (e.g. the trailing "улс ор" in the first row, or "өхөрсөн" missing its leading "с" from "сөхөрсөн"). The ASR transcript is generated per segment, so it never shows these edge artifacts.
- Edited records. Official parliamentary records are lightly edited, fillers ("аа") and false starts are removed, and typos occur, while the ASR output is verbatim. So at segment edges the ASR text can indeed look locally more accurate, but mid-segment the human transcript is the reliable reference. For stricter subsets we recommend filtering on the cer column.