whisper-small-zh-TW / README.md
xmzhu's picture
update model card README.md
d75cc79
metadata
language:
  - zh
license: apache-2.0
tags:
  - whisper-event
  - generated_from_trainer
datasets:
  - mozilla-foundation/common_voice_11_0
metrics:
  - wer
model-index:
  - name: Whisper Small Chinese (Taiwanese Mandarin)
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: mozilla-foundation/common_voice_11_0 zh-TW
          type: mozilla-foundation/common_voice_11_0
          config: zh-TW
          split: test
          args: zh-TW
        metrics:
          - name: Wer
            type: wer
            value: 42.988741044012286

Whisper Small Chinese (Taiwanese Mandarin)

This model is a fine-tuned version of xmzhu/whisper-small-zh on the mozilla-foundation/common_voice_11_0 zh-TW dataset. It achieves the following results on the evaluation set:

  • Loss: 0.2639
  • Wer: 42.9887

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 1e-05
  • train_batch_size: 64
  • eval_batch_size: 32
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • training_steps: 5000
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
0.0058 6.02 1000 0.2445 43.0911
0.0006 13.02 2000 0.2639 42.9887
0.0003 20.01 3000 0.2787 43.1934
0.0002 27.0 4000 0.2877 43.5415
0.0002 33.02 5000 0.2910 43.5824

Framework versions

  • Transformers 4.26.0.dev0
  • Pytorch 1.13.1+cu117
  • Datasets 2.7.1.dev0
  • Tokenizers 0.13.2