jethrowang/whisper-tiny_hat-esc_vanilla

Whisper Tiny Hakka (vanilla)

This model is a fine-tuned version of openai/whisper-tiny on the HAT ASR Aligned dataset. It achieves the following results on the evaluation set:

• Loss: 0.2624
• Cer: 394.3662

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: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 488
• training_steps: 4880
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
0.5947	0.9980	488	0.8494	163.1077
0.1948	1.9959	976	0.4130	139.9531
0.1153	2.9939	1464	0.3256	138.9243
0.0806	3.9918	1952	0.2937	158.7396
0.0561	4.9898	2440	0.2781	235.7249
0.0418	5.9877	2928	0.2671	289.3289
0.0294	6.9857	3416	0.2672	294.8922
0.025	7.9836	3904	0.2618	324.0955
0.0214	8.9816	4392	0.2617	363.4556
0.0159	9.9796	4880	0.2624	394.3662

Framework versions

• Transformers 4.49.0
• Pytorch 2.3.0
• Datasets 3.3.2
• Tokenizers 0.21.0