whisper-medium-zh-20230712 - au2a

This model is a fine-tuned version of openai/whisper-medium on the some hakka audio dataset. It achieves the following results on the evaluation set:

• Loss: 0.2659
• Cer: 87.6898

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: 5e-06
• train_batch_size: 8
• eval_batch_size: 8
• 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: 15000

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
0.2417	0.16	1000	0.3919	92.1659
0.1219	0.32	2000	0.2963	81.3855
0.0762	0.49	3000	0.2785	68.9544
0.0524	0.65	4000	0.2660	89.4916
0.0347	0.81	5000	0.2517	96.8800
0.0255	0.97	6000	0.2567	89.0232
0.0104	1.13	7000	0.2547	91.9959
0.0069	1.29	8000	0.2609	85.5481
0.0072	1.46	9000	0.2605	72.8148
0.0081	1.62	10000	0.2593	81.8161
0.0024	1.78	11000	0.2608	79.6064
0.0021	1.94	12000	0.2622	78.2655
0.0004	2.1	13000	0.2656	86.0580
0.0005	2.27	14000	0.2665	90.1677
0.0005	2.43	15000	0.2659	87.6898

Framework versions

• Transformers 4.30.2
• Pytorch 1.11.0+cu113
• Datasets 2.13.1
• Tokenizers 0.13.3