Datasets:

---

FBK-MT

/

mosel

like 85

Follow

FBK-MT 35

Is this update applied to a selected subset?
Also, does the updated text remain paired with the VoxPopuli-split audio, or is it now aligned with the different audio segments extracted by NeMo?

This update is applied to all subsets. You can still access the previous version of the dataset using v1.1 (https://huggingface.co/datasets/FBK-MT/mosel/tree/v1.1). The new texts are aligned with the same VoxPopuli split audio.

Thanks for clarifying.
I have one more question regarding the NeMo-speech-data-processor update — did you apply VAD and run inference on the individual speech segments (with possible transcript filtering at the segment level), or were the VAD-detected speech parts concatenated into a single file before inference?

I’m asking because I noticed that in some cases, such as this example:
audio id 20160118-0900-PLENARY-10_pt_0,
the updated transcript (v2) misses the first word “artigo”, which is still present in both version 1 and the corresponding audio sample.

Could this difference be related to how VAD or segment-level inference was handled in the SDP pipeline?

Please direct this question to the NeMo repository or the granary HF repo, as they can give better answer than me.

Thanks!