whisper-large-v3-ft-btb-cv-cvad-ca-cy-2511

This model is a fine-tuned version of openai/whisper-large-v3 on the techiaith/banc-trawsgrifiadau-bangor train 25.10, techiaith/commonvoice_23_0_cy train+dev+other_with_excluded main, cymen-arfor/lleisiau-arfor train+dev main, techiaith/commonvoice_vad_cy train main dataset. It achieves the following results on the evaluation set:

• Loss: 0.3775
• Wer: 0.2938

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
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 6000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4961	0.5928	500	0.5267	0.3950
0.3376	1.1849	1000	0.4401	0.3313
0.3166	1.7777	1500	0.3996	0.3110
0.2334	2.3699	2000	0.3908	0.2970
0.2258	2.9627	2500	0.3775	0.2938
0.1559	3.5548	3000	0.3949	0.2881
0.0973	4.1470	3500	0.4325	0.2875
0.0986	4.7398	4000	0.4370	0.2843
0.0578	5.3320	4500	0.4849	0.2897
0.0568	5.9247	5000	0.4883	0.2915
0.0377	6.5169	5500	0.5269	0.2902
0.0348	7.1091	6000	0.5282	0.2905

Framework versions

• Transformers 4.52.4
• Pytorch 2.9.0+cu128
• Datasets 4.4.0
• Tokenizers 0.21.4