whisper-base-zh-20230721 - au2a

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

• Loss: 0.4546
• Cer: 16.5974

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: 32
• 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: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
0.4669	0.65	1000	0.6528	25.0548
0.2208	1.29	2000	0.5006	19.8761
0.1452	1.94	3000	0.4546	17.9497
0.0951	2.59	4000	0.4431	17.4511
0.0526	3.24	5000	0.4450	17.3113
0.0422	3.88	6000	0.4440	16.6201
0.0271	4.53	7000	0.4471	17.0658
0.0179	5.18	8000	0.4509	16.5823
0.0166	5.83	9000	0.4535	16.8543
0.0129	6.47	10000	0.4546	16.5974

Framework versions

• Transformers 4.30.2
• Pytorch 2.0.1+cu117
• Datasets 2.13.1
• Tokenizers 0.13.3