This repository contains the archive of CV10 (test set) with checked Ukrainian transcriptions and audios. All audios have been checked by a human to be sure that they are correct.
This archive is used to test all ASR models listed here: https://github.com/egorsmkv/speech-recognition-uk
Example with datasets:
from datasets import load_dataset
ds = load_dataset('Yehor/cv10-uk-testset-clean')
print(ds)
for row in ds['train']:
audio = row["audio"]
sampling_rate = audio["sampling_rate"]
audio_bytes = audio["array"]
filename = audio["path"]
print(len(audio_bytes), sampling_rate, filename)
print(row["duration"], row["transcription"])
print('---')Example with polars: https://colab.research.google.com/drive/1upeXw3WbLjK37b1LetpM0HxFXDdOZqSK?usp=sharing
Use the following colabs to see how you can download this dataset in Python:
datasets:
polars:
Duration: 4.6 hours
| Metrics | Value |
|---|---|
| mean | 5.201474 |
| std | 1.764957 |
| min | 1.704 |
| 25% | 3.816 |
| 50% | 4.896 |
| 75% | 6.384 |
| max | 10.536 |
We recommend to use Hugging Face dataset, but in case you need raw dataset, use:
-
Audio data: https://github.com/egorsmkv/cv10-uk-testset-clean/releases/download/v1.1/filtered-cv10-test.zip
-
Labels list (TAB format) with absolute paths: https://github.com/egorsmkv/cv10-uk-testset-clean/blob/main/labels_absolute.lst
-
Labels list (CSV format) with absolute paths: https://github.com/egorsmkv/cv10-uk-testset-clean/blob/main/labels_absolute.csv
-
Labels list (CSV format) with relative paths: https://github.com/egorsmkv/cv10-uk-testset-clean/blob/main/labels_relative.csv