GigaSpeech on HuggingFace

[dophist]

GigaSpeech dataset is now available on HuggingFace Hub.

Highlights of GigaSpeech on HuggingFace

• easy to use (a two-liner in python)
• Smoother and faster downloading from US & EU, even support on-the-fly downloading during training
• preprocessed:
  • decompressed
  • short audio files(.wav) are segmented and extracted from raw long audio
  • supervisions are extracted from raw metadata.json
• subsets can be downloaded separately (e.g. XS/S/M/L/XL for training, DEV/TEST for benchmarking)
• users can even listen to audio samples via HuggingFace's dataset viewer

How-to

• Step 1: Fill GigaSpeech application form from SpeechColab
• Step 2: Go to GigaSpeech on HuggingFace webpage and follow instructions there.

Useful links

• twitter announcement
• GigaSpeech's main page on HuggingFace

Credits

Many thanks to The Dataset Team & Speech Team at HuggingFace, particularly @polinaeterna , @patrickvonplaten , @sanchit-gandhi , GigaSpeech just becomes more accessible to the entire speech community!