Jingmiao/whisper-small-zh_tw

Whisper Small Chinese (Taiwan)

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

• Loss: 0.2283
• Wer: 41.9652

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
0.0049	6.02	1000	0.2283	41.9652
0.0008	13.02	2000	0.2556	42.0266
0.0004	20.01	3000	0.2690	42.4156
0.0003	27.0	4000	0.2788	42.7840
0.0002	33.02	5000	0.2826	43.0297

Framework versions

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