mpasila/faster-whisper-large-finnish-v3

This is a conversion of Finnish-NLP/whisper-large-finnish-v3 into faster-whisper format.

This is our improved Whisper v3 model that is now finetuned from OpenAI Whisper Large V3

We improve from our previously finetuned Whisper V2 model in the following mannerhttps://huggingface.co/Finnish-NLP/whisper-large-v2-finnish

CV11 (Common Voice 11 test set) WER (Word error rate) 10.42 --> 8.23

Fleurs (A speech recognition test set by Google) WER (Word error rate) 10.20 --> 8.21

Model was trained on Nvidia RTX4080 for 32k steps with batch size 8, gradient accumulation 2

Original OpenAI Whisper Large V3

- CV11 - WER: 14.81 - WER NORMALIZED: 10.82 - CER: 2.7 - CER NORMALIZED: 2.07

• Fleurs
  • WER: 12.04
  • WER NORMALIZED: 9.63
  • CER: 2.48
  • CER NORMALIZED: 3.64

After Finetuning with Finnish data our V3 got these scores on the test set:

• @14000 finetuning steps

  • CV11

    • WER: 11.36
    • WER NORMALIZED: 8.31
    • CER: 1.93
    • CER NORMALIZED: 1.48

  • Fleurs

    • WER: 10.2
    • WER NORMALIZED: 8.56
    • CER: 2.26
    • CER NORMALIZED: 3.54

• @32000 finetuning steps

  • CV11

    • WER: 11.47
    • WER NORMALIZED: 8.23
    • CER: 1.91
    • CER NORMALIZED: 1.43

  • Fleurs

    • WER: 10.1
    • WER NORMALIZED: 8.21
    • CER: 2.2
    • CER NORMALIZED: 3.23