Add train and test splits

audeering / emodb

Publishes Berlin Database of Emotional Speech with audb

Other

2 stars 0 forks source link

Add train and test splits #4

Closed audeerington closed 2 years ago

audeerington commented 2 years ago

Adds unofficial, speaker-independent train and test splits in the tables emotion.categories.train.gold_standard and emotion.categories.test.gold_standard.

The speakers are distributed like this:

train: [3, 8, 9, 10, 11, 13] (3 female, 3 male)
test: [12, 14, 15, 16] (2 female, 2 male)

These are the resulting histograms:

train_gender_white_bg test_gender_white_bg

The table emotion is not removed to keep this version compatible with examples and tutorials, since emodb is often used as an example.

hagenw commented 2 years ago

Seems like females are the better actors. In the emodb paper it is stated that each actor acted the same sentences and we had 800 sentences altogether, 400 for females and 400 for males. Afterwards the emotions were annotated and samples with low recognition rate were removed leaving 302 female and 233 male samples.

Your changes look fine and it makes indeed sense to leave the emotion table as emodb is special as it is used in demos and even unit tests.