xezpeleta/whisper-large-eu

Whisper Large Basque

This model is a fine-tuned version of openai/whisper-large-v3 on the mozilla-foundation/common_voice_17_0 eu dataset. It achieves the following results on the evaluation set:

• Loss: 0.1259
• Wer: 7.2154

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: 4.375e-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: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2208	0.05	500	0.2592	20.6915
0.1489	0.1	1000	0.1971	14.6827
0.1973	0.15	1500	0.1747	12.3777
0.1353	1.0296	2000	0.1527	10.7195
0.1065	1.0796	2500	0.1456	9.8694
0.106	1.1296	3000	0.1362	9.0925
0.0718	2.0092	3500	0.1326	8.5428
0.0683	2.0592	4000	0.1343	8.4851
0.0482	2.1092	4500	0.1336	8.1049
0.0548	2.1592	5000	0.1316	7.9244
0.0282	3.0388	5500	0.1391	7.8182
0.025	3.0888	6000	0.1425	7.9409
0.0274	3.1388	6500	0.1391	7.7311
0.0155	4.0184	7000	0.1492	7.6972
0.0189	4.0684	7500	0.1517	7.6569
0.0139	4.1184	8000	0.1539	7.6267
0.0141	4.1684	8500	0.1550	7.5424
0.0368	5.048	9000	0.1259	7.2154

Framework versions

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