Datasets:
Transcript quality issue in Vietnamese dataset
Hi,
Thank you for releasing this dataset. I am opening this issue because I noticed that the Vietnamese transcripts appear to be incorrect across the dataset.
From the samples I checked, the transcripts do not seem to match the corresponding audio. This does not look like a small number of annotation mistakes, but rather a systematic issue. It may indicate that something went wrong in the data processing or transcript-generation pipeline, such as misaligned audio/text pairs, incorrect mapping between files and transcripts, or an ASR/transcription step being applied incorrectly.
Could you please re-check the Vietnamese dataset and verify whether the transcripts are correctly aligned with the audio files?
This issue is important because the current transcripts may make the dataset difficult or unsafe to use for training or evaluation.
Thanks again for your work on the dataset.
Hi, thank you very much for pointing this out, we're looking into it!
Hi @khangnguyen2 , following up on this, thanks again for your comment,
We re-checked the vn_vi config by sampling segments across the whole dataset and could not reproduce a systematic audio–transcript mismatch: human_transcript and asr_transcript agree closely throughout (mean CER ≈ 0.15, and our alignment filter caps it at 0.30).
Since the ASR transcript is generated directly from each audio segment, this agreement would be impossible if the audio and human transcripts were systematically misaligned. One thing to double-check on your side: human_transcript is the reference text; asr_transcript is raw ASR output and does contain phonetic errors (e.g. "mai-nici sim-bun" for "Mainichi Shimbun"), if you were reading that column, it would indeed look wrong. If you are seeing mismatches in the human_transcript column, could you please share a few segment_ids?
We'd be happy to investigate further examples! Any resulting fixes to the dataset would land after the 24th of September.