Skip to content

Dataset · alvanlii

10k hours of Cantonese YouTube speech

alvanlii/cantonese-youtube

Large-scale corpus of 10,000 hours of Cantonese YouTube audio with pseudo-transcriptions from Whisper and SenseVoice.

What's inside

The dataset contains between 1 and 10 million records with a total duration of about 10,000 hours in Cantonese and Chinese. Audio is sourced from random YouTube channels across various topics: news, vlogs, entertainment, stories, and health. Instead of manual transcription, the columns contain automated transcripts (in Traditional Chinese) generated by the whisper-large-v2-cantonese and SenseVoiceSmall models.

What you can do with it

  • Large-scale pre-training of ASR models for the Cantonese dialect
  • Training audio classifiers by content topic
Tasks
Automatic speech recognition, Audio classification
Languages
Chinese, Cantonese
Size
1M–10M rows
Downloads this month
296
Likes
50
Access
on request on Hugging Face

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("alvanlii/cantonese-youtube")