Datasets:
language:
- cy
license: cc0-1.0
size_categories:
- 1K<n<10K
task_categories:
- automatic-speech-recognition
dataset_info:
features:
- name: sentence
dtype: string
- name: audio
dtype:
audio:
sampling_rate: 16000
splits:
- name: dev
num_bytes: 141550478
num_examples: 260
- name: test
num_bytes: 131678956
num_examples: 235
- name: train
num_bytes: 1268995249.832
num_examples: 2339
download_size: 1372623929
dataset_size: 1542224683.832
configs:
- config_name: default
data_files:
- split: dev
path: data/dev-*
- split: test
path: data/test-*
- split: train
path: data/train-*
The training and development set sentences are taken from CoVoST and have been compared to all validated sentences in the Welsh Common Voice data to ensure none of the already recorded sentences will be used here. Then all sentences containing personal names have been extracted and replaced with a randomly generated name using the Faker library and a custom Welsh names list. The sentences were then recorded by 26 volunteers from North-West Wales, 15 women, 10 men and one non-binary person.
The test split of this dataset is based on the Welsh Common Voice test split. All sentences containing a personal name have been selected and the names have been replaced by 64 Welsh first names each occurring 6 times in combination with either another first name (e.g. the parents first name according to patronymic or matronymic tradition or a second first name to help distinguish oneself if the last name is very common), the word ap or ferch (meaning son or daughter of) + the another first name or with one of 13 common Welsh last names. They were also recorded by the same 26 volunteers.
Similarly to Common Voice, the participants were able to record the sentences by themselves using their own hardware which resulted in a variety of sound qualities. Speakers have been given a number and can, if needed, be excluded from the data.
All splits together add up to 4 Hours : 27 Minutes : 50 Seconds