10,000 hours of Cantonese audio sourced from YouTube across categories including news, vlogs, entertainment, stories, and health. It features dual pseudo-transcriptions generated by Whisper-large-v2 and SenseVoiceSmall, both converted to Traditional Chinese using OpenCC.
Use Cases
- Benchmark ASR model performance by comparing transcript_whisper against transcript_sensevoice.
- Train Cantonese-specific language models using the large-scale text data in the transcript columns.
- Conduct topic modeling or linguistic analysis across the news, vlogs, and health categories.
Strengths
- 10,000 hours of audio data scraped from diverse YouTube channels.
- Dual transcription columns: transcript_whisper and transcript_sensevoice.
- Content spans five distinct categories: news, vlogs, entertainment, stories, and health.
- Transcriptions are normalized to Traditional Chinese using OpenCC.