xezpeleta/whisper-medium-eu

Whisper Medium Basque

This model is a fine-tuned version of openai/whisper-medium on the mozilla-foundation/common_voice_17_0 eu dataset. It achieves the following results on the evaluation set:

• Loss: 0.1787
• Wer: 8.8021

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
0.3171	0.0625	500	0.3369	25.5304
0.1852	0.125	1000	0.2409	17.3110
0.2353	0.1875	1500	0.2050	14.2228
0.1569	1.037	2000	0.1815	12.2861
0.125	1.0995	2500	0.1692	11.1144
0.12	1.162	3000	0.1600	10.6975
0.069	2.0115	3500	0.1540	9.7649
0.0606	2.074	4000	0.1550	9.8199
0.0434	2.1365	4500	0.1580	9.4571
0.0455	2.199	5000	0.1533	9.1410
0.0216	3.0485	5500	0.1620	9.0842
0.017	3.111	6000	0.1704	9.0980
0.0174	3.1735	6500	0.1681	9.0723
0.0098	4.023	7000	0.1725	8.8625
0.0076	4.0855	7500	0.1765	8.8351
0.007	4.148	8000	0.1787	8.8021

Framework versions

• Transformers 4.46.0.dev0
• Pytorch 2.4.1+cu121
• Datasets 3.0.2.dev0
• Tokenizers 0.20.0