zuazo/whisper-large-eu-train

Whisper Large Basque

This model is a fine-tuned version of openai/whisper-large on the mozilla-foundation/common_voice_13_0 eu dataset. It achieves the following results on the evaluation set:

• Loss: 0.4229
• Wer: 13.1677

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: 32
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 20000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.067	5.85	1000	0.2644	15.8677
0.0123	11.7	2000	0.3077	14.6326
0.0052	17.54	3000	0.3317	14.1853
0.0037	23.39	4000	0.3387	14.0885
0.0026	29.24	5000	0.3559	14.2618
0.0026	35.09	6000	0.3604	14.2155
0.002	40.94	7000	0.3734	14.1228
0.0012	46.78	8000	0.3773	14.0301
0.0012	52.63	9000	0.3802	13.9072
0.0012	58.48	10000	0.3850	14.4734
0.0006	64.33	11000	0.3896	13.6513
0.0011	70.18	12000	0.3981	13.6311
0.001	76.02	13000	0.3947	13.5949
0.0002	81.87	14000	0.4039	13.6170
0.0001	87.72	15000	0.4057	13.4579
0.0008	93.57	16000	0.4119	13.2745
0.0001	99.42	17000	0.4203	13.1717
0.0001	105.26	18000	0.4166	13.0972
0.0001	111.11	19000	0.4243	13.0448
0.0	116.96	20000	0.4229	13.1677

Framework versions

• Transformers 4.37.2
• Pytorch 2.2.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1