xezpeleta/whisper-medium-eu

Whisper Medium Basque

This model is a fine-tuned version of openai/whisper-medium on the asierhv/composite_corpus_eu_v2.1 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2191
• Wer: 9.9841
• Wer on Common Voice 17.0, test split: 7.97

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: 6.25e-06
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.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: 500
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3412	0.0625	500	0.4570	28.2182
0.1462	0.125	1000	0.3524	19.9402
0.2495	0.1875	1500	0.3077	17.5236
0.2617	0.25	2000	0.2811	16.5841
0.1646	0.3125	2500	0.2726	13.8917
0.0934	0.375	3000	0.2533	14.0273
0.1016	0.4375	3500	0.2331	12.1623
0.1454	0.5	4000	0.2299	11.5546
0.1502	0.5625	4500	0.2333	12.4007
0.0916	0.625	5000	0.2271	10.9657
0.0914	0.6875	5500	0.2343	10.5029
0.1093	0.75	6000	0.2191	9.9841
0.0948	0.8125	6500	0.2215	10.5357
0.0744	0.875	7000	0.2108	11.2368
0.1269	0.9375	7500	0.2158	10.0028
0.1408	1.0	8000	0.2141	10.1290

Framework versions

• Transformers 4.49.0.dev0
• Pytorch 2.6.0+cu124
• Datasets 3.2.1.dev0
• Tokenizers 0.21.0