Datasets:
metadata
tags:
- speech
- speech-transcription
- romanian
language:
- ro
license: mit
Common Voices Corpus 20.0 (Romanian)
Common Voices is an open-source dataset of speech recordings created by Mozilla to improve speech recognition technologies. It consists of crowdsourced voice samples in multiple languages, contributed by volunteers worldwide.
Challenges: The raw dataset included numerous recordings with incorrect transcriptions or those requiring adjustments, such as sampling rate modifications, conversion to .wav format, and other refinements essential for optimal use in developing and fine-tuning various models.
Processing: Our team, led by project manager Vișan Ionuț, carefully reviewed and manually corrected the transcriptions of all audio segments, ensuring their conversion into the required format for modern models (16k Hz sampling rate, mono channel, .wav format).
Dataset Summary
common_voices20_audio: The folder containing all processed audio segments.
Total number of audio segments: 41,431.
Total duration of all audio segments combined: approximately 47 hours.
common_voices20.csv: Contains metadata for all segments from the common_voices20_audio.
The file contains 41,431 rows and 2 columns:
- audio_file: File names of the processed audio segments from common_voices20_audio.
- transcript: Corresponding text transcriptions for each audio file from common_voices20_audio.
Split
To split the dataset (common_voices20.csv), we performed an 80-20 split into training and test sets using a seed value of 42, resulting in:
- train_common_voices20.csv: It contains 33,144 of the audio segments.
- test_common_voices20.csv: It contains 8,287 of the audio segments.
How to use
from datasets import load_dataset
dataset = "TransferRapid/CommonVoices20_ro"
dataset = load_dataset(dataset)
print(dataset)
from datasets import load_dataset
dataset = "TransferRapid/CommonVoices20_ro"
dataset = load_dataset(dataset)
print(dataset)