Datasets:

---

google

/

WaxalNLP

like 203

Follow

Google 46.9k

Hello, and thank you for releasing the WaxalNLP dataset, this is a very valuable resource.

I am currently working with the ASR subsets (ach_asr) and wanted to clarify something regarding the dataset splits.

For each language individually, I:

1. Loaded the train, validation, and test splits.
2. Extracted the speaker_id field from each split.
3. Constructed sets of unique speaker IDs per split.
4. Computed intersections between the splits to check for speaker overlap.

For example (ach_asr yields the results below):

• Train speakers: 322
• Validation speakers: 199
• Test speakers: 194

Overlap counts:

• Train ∩ Validation: 194
• Train ∩ Test: 192
• Validation ∩ Test: 150

Based on this, it appears that many speakers occur in multiple splits.

I wanted to ask:

• Is speaker overlap across splits intentional?
• Were the splits designed at the utterance level rather than the speaker level?
• For speaker-independent ASR evaluation, would it be appropriate to re-split the data by speaker_id?

I just want to make sure I understand the intended usage and evaluation protocol before proceeding.

Thank you again for making this dataset available.

EDIT: You can check out the script I am working with here: https://colab.research.google.com/drive/1Zpvjl3DBBEi22HTqbUcaEZWvcA0KmJRk?usp=sharing