Datasets:
Extra Words in Audio Not Reflected in Transcript
I found a potential issue in the dataset. For samples where the transcript is not a complete sentence, the corresponding audio sometimes contains additional words at the end that are not present in the transcript.
Example:
dataset_part_070.parquet
Row: 3
Audio: 703-4.wav
This may indicate that the transcript was truncated during preprocessing while the original audio was left unchanged, or that the sentence segmentation between the audio and transcript became misaligned. Such mismatches could introduce noise during training, especially for TTS or speech-text alignment models.
It might be worth checking whether this affects other samples in the dataset as well.
Thank you for reporting this issue, Mr. Abbas Zadeh. As discussed in the paper related to this dataset, the transcripts are not a perfect match to the audio all the time. There is an estimated character error rate of about 1%. However, this issue can help other developers filter out the less certain samples for more sensitive tasks. I appreciate you bringing this to our attention.