alvanlii/cantonese-youtube

Cantonese Youtube Pseudo-Transcription Dataset

• Contains approximately 10k hours of audio sourced from YouTube
  • Videos are chosen at random, and scraped on a channel basis
  • Includes news, vlogs, entertainment, stories, health
• Columns
  • transcript_whisper: Transcribed using Scrya/whisper-large-v2-cantonese with alvanlii/whisper-small-cantonese for speculative decoding
  • transcript_sensevoice: Transcribed using FunAudioLLM/SenseVoiceSmall
    • used OpenCC to convert to traditional chinese
    • isolated event tags to event_sensevoice
    • isolated emotion tags to emotion_sensevoice
  • snr: Signal-to-noise ratio, extracted from ylacombe/brouhaha-best
  • c50: Speech clarity, extracted from ylacombe/brouhaha-best
  • emotion: Emotion, extracted from emotion2vec/emotion2vec_plus_large
  • Note that id does not reflect the ordering of the audio within the same video
• Processing
  • The full audio is split using WhisperX, using Scrya/whisper-large-v2-cantonese
    • it is split in <30s chunks and according to speakers
  • Preliminary filtering includes filtering out phrases like:
    • like/subscribe to YouTube channel
    • subtitles by [xxxx]
    • Additional filtering is recommended for your own use
• Note: An earlier version of the dataset has duplicated data. I recommend re-downloading it if you downloaded it before Nov-7-2024