Kushtrim/whisper-large-v3-turbo-shqip-50h

Whisper Large v3 Turbo Shqip

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Audio Shqip 50 orë dataset. It achieves the following results on the evaluation set:

• Loss: 0.5501
• Wer: 26.2952

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• 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
0.5589	0.5363	500	0.5765	40.2773
0.391	1.0727	1000	0.4633	32.9234
0.3557	1.6090	1500	0.4209	32.8188
0.2288	2.1453	2000	0.4132	30.0056
0.237	2.6817	2500	0.4012	29.9073
0.1776	3.2180	3000	0.4055	30.2650
0.1838	3.7544	3500	0.4034	29.6501
0.1328	4.2907	4000	0.4109	29.3719
0.1301	4.8270	4500	0.4052	28.7716
0.1034	5.3634	5000	0.4231	27.3180
0.0845	5.8997	5500	0.4296	27.5167
0.0857	6.4360	6000	0.4526	26.9750
0.0526	6.9724	6500	0.4550	27.2343
0.0436	7.5087	7000	0.4833	27.2824
0.0284	8.0451	7500	0.4983	26.5734
0.0328	8.5814	8000	0.5043	26.8244
0.0164	9.1177	8500	0.5225	26.5441
0.0171	9.6541	9000	0.5318	26.2659
0.019	10.1904	9500	0.5473	26.3182
0.0253	10.7267	10000	0.5501	26.2952

Framework versions

• Transformers 4.45.2
• Pytorch 2.5.1+cu121
• Datasets 3.1.0
• Tokenizers 0.20.3