metadata
dataset_info:
features:
- name: audio
dtype: audio
- name: speaker
dtype: string
- name: transcription
dtype: string
- name: translation
dtype: string
- name: Class
dtype: string
- name: Class_index
dtype: float64
splits:
- name: train
num_bytes: 515094059
num_examples: 750
- name: test
num_bytes: 171698025
num_examples: 250
download_size: 590736573
dataset_size: 686792084
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
license: cc-by-nc-nd-4.0
language:
- it
- en
concatentation of "thomasavare/waste-classification-audio-deepl" and "thomasavare/waste-classification-audio-deepl2", so its easier to use.
there are 3 duplicates, idk why but it's too long to remake the datasets (maybe later) and not necessary.