Datasets:

---

laion

/

Emolia

like 8

Follow

LAION eV 738

Dear Laion team,

I would like to use this dataset for a speech model project. However, I have come across a few questions that I think are relevant to ask the community:

1.

You say you use Orange's Speaker-wavLM-tbr model, but this model only allows you to produce embeddings corresponding to each speaker's timbre.
So I was wondering how you generated these annotations: age, gender, dominance, expressiveness, etc.

My hypotheses:

1. You used a model that you did not mention in README.md.
2. For each label, you designated an audio clip corresponding to the total exaggeration of the trait:
   for example, to calculate the humor score, you took a very humorous audio clip and calculated the cosine similarity with each sample.

2.

In the emotion annotation category, many labels are not explained in the README (their definition range and how to interpret them).
For example:
Amusement_Best
Confident_vs._Hesitant_best (How should this even be interpreted?)
Contemplation_best...

I find that very ambiguous...

Thanks for answering these questions ❤️