whisper-large-v3-ft-cy-en

This model is a version of openai/whisper-large-v3 fine-tuned with a curated collection of Welsh and English speech data (see: techiaith/commonvoice_18_0_cy_en collected originally from Mozilla's Common Voice project.

It achieves the following results on the following language specific test sets:

• WER (test_en): 13.85
• WER (test_cy): 8.78
• WER (test_cy+test_en): 9.55

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2097	0.2497	1000	0.2169	14.2221
0.1621	0.4993	2000	0.1816	11.6845
0.1406	0.7490	3000	0.1609	10.2445
0.1242	0.9987	4000	0.1505	9.5594

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1