zuazo/whisper-large-gl-train

Whisper Large Galician

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

• Loss: 0.3219
• Wer: 6.5005

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
0.0476	5.83	1000	0.1829	6.3535
0.008	11.66	2000	0.2224	6.2705
0.0043	17.49	3000	0.2360	6.3397
0.0029	23.32	4000	0.2544	6.5386
0.0036	29.15	5000	0.2552	6.6977
0.0026	34.99	6000	0.2737	6.8568
0.0009	40.82	7000	0.2734	6.6320
0.0009	46.65	8000	0.2769	6.8187
0.0006	52.48	9000	0.2832	6.6164
0.0013	58.31	10000	0.2883	7.0176
0.0005	64.14	11000	0.2972	6.8983
0.0006	69.97	12000	0.2964	6.6735
0.0003	75.8	13000	0.3042	6.7392
0.0002	81.63	14000	0.3084	6.7426
0.0001	87.46	15000	0.3145	6.6631
0.0002	93.29	16000	0.3091	6.6666
0.0001	99.13	17000	0.3170	6.8758
0.0002	104.96	18000	0.3223	6.6337
0.0	110.79	19000	0.3219	6.4971
0.0001	116.62	20000	0.3219	6.5005

Framework versions

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