metadata

license: apache-2.0
tags:
  - whisper-medium
  - asr
  - zh-TW
datasets:
  - mozilla-foundation/common_voice_11_0
model-index:
  - name: Whisper Medium TW
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        dataset:
          name: mozilla-foundation/common_voice_11_0
          type: mozilla-foundation/common_voice_11_0
          config: zh-TW
          split: test
        metrics:
          - type: wer
            value: 7.38
            name: WER

Whisper Medium TW

This model is a fine-tuned version of openai/whisper-medium on the mozilla-foundation/common_voice_11_0 dataset.

Training and evaluation data

Training:

mozilla-foundation/common_voice_11_0 (train+validation)

Evaluation:

mozilla-foundation/common_voice_11_0 (test)

Training procedure

Datasets were augmented using audiomentations via PitchShift, TimeStretch, Gain, AddGaussianNoise transformations at p=0.3.
A space is added between each Chinese character, as demonstrated in the original paper. Effectively, WER == CER in this case.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 1
eval_batch_size: 1
gradient_accumulation_steps: 32
optimizer: Adam
generation_max_length: 225,
warmup_steps: 200
max_steps: 2000,
fp16: True,
evaluation_strategy: "steps",

Framework versions

Transformers 4.27.1
Pytorch 2.0.1+cu120
Datasets 2.13.1