whisper-medium-ctc-salt-v4

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3685
• Wer: 0.4950
• Cer: 0.1125

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: 3e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.1
• training_steps: 7500

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.8100	0.0692	500	1.4483	0.9447	0.3121
1.1358	0.1384	1000	0.7928	0.8266	0.2115
0.9544	0.2076	1500	0.6407	0.7104	0.1784
0.9246	0.2768	2000	0.5626	0.6579	0.1615
0.8689	0.3460	2500	0.5141	0.6114	0.1470
0.8005	0.4152	3000	0.4850	0.5991	0.1431
0.7604	0.4844	3500	0.4607	0.5729	0.1341
0.7502	0.5536	4000	0.4346	0.5531	0.1290
0.7427	0.6228	4500	0.4137	0.5387	0.1247
0.7206	0.6920	5000	0.4008	0.5212	0.1189
0.6932	0.7612	5500	0.3868	0.5056	0.1156
0.7040	0.8304	6000	0.3777	0.4989	0.1136
0.6987	0.8997	6500	0.3729	0.5001	0.1140
0.6623	0.9689	7000	0.3692	0.4945	0.1122
0.6484	1.0381	7500	0.3685	0.4950	0.1125

Framework versions

• Transformers 5.2.0
• Pytorch 2.10.0+cu130
• Datasets 4.6.0
• Tokenizers 0.22.2