ghost613's picture
Update README.md
3ff2d57 verified
|
raw
history blame
796 Bytes
metadata
datasets:
  - Bingsu/zeroth-korean
language:
  - ko
metrics:
  - cer
  - wer
base_model:
  - openai/whisper-large-v3-turbo
pipeline_tag: automatic-speech-recognition

Description

Fine-tuning Whisper Large V3 Turbo on zeroth Korean dataset.

Dataset split:

  • The test dataset from Korean zeroth is divided to test and validation -> 50% validation, 50% test
  • Train set duration: 206 hours 43 minutes
  • Validation set duration: 2 hours 22 minutes
  • Test set duration: 2 hours 22 minutes

Results:

  • initial validation WER: 26.26%
  • final validation WER: 4.90%
  • initial validation CER: 6.67%
  • final validation CER: 1.78%
  • initial test WER: 26.75%
  • final test WER: 4.89%
  • initial test CER: 7.58%
  • final test CER: 2.06%

Notes

  • Models did not converge, better results are possible.