Datasets:

---

FBK-MT

/

mosel

like 78

Follow

FBK-MT 31

Hello,

I'm working with the MOSEL v2 dataset and trying to process LibriLight audio files to match the TSV annotations. I've noticed that the TSV entries reference pre-segmented audio files (e.g., 3483_2938_thebeetle_21_marsh_64kb_0007) with specific offset and duration values.

However, I'm facing challenges:

1. The original LibriLight dataset contains long audio files, not the segmented versions referenced in the TSV
2. Facebook's pre-segmented data appears to be inaccessible, and while their segmentation tools are available, they don't reproduce the exact segments referenced in the TSV files
3. The TSV segment IDs (like _0007) don't correspond to actual files in the LibriLight distribution

Questions:

• Is there an official source for the pre-segmented LibriLight audio files that match the MOSEL TSV entries?
• If not, could you provide the segmentation boundaries or voice activity detection (VAD) information used to create these segments?

Request:
If the pre-segmented data is unavailable, would it be possible to share:

• The specific segmentation method or parameters used (e.g., VAD thresholds, segment duration targets)
• A mapping between TSV segment IDs and the original audio files with precise timestamps
• Or any guidance on how to accurately reproduce the segmentation

This information would greatly help ensure proper alignment between audio and text annotations in the dataset.

Thank you for your excellent work on this dataset.