Kushtrim/whisper-large-v3-turbo-shqip-115h

Whisper Large v3 Turbo Shqip

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Audio Shqip 115 orë dataset. It achieves the following results on the evaluation set:

• Loss: 0.3322
• Wer: 22.0069

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
0.5211	0.2738	500	0.5221	36.9257
0.4152	0.5476	1000	0.4144	31.1469
0.3847	0.8215	1500	0.3747	28.2953
0.2703	1.0953	2000	0.3536	26.4348
0.2471	1.3691	2500	0.3419	25.5897
0.2691	1.6429	3000	0.3293	24.5533
0.2426	1.9168	3500	0.3202	24.5742
0.1993	2.1906	4000	0.3178	23.5548
0.204	2.4644	4500	0.3124	23.6609
0.2	2.7382	5000	0.3098	23.5131
0.1298	3.0120	5500	0.3101	22.5753
0.1213	3.2859	6000	0.3145	23.0129
0.1343	3.5597	6500	0.3105	22.6511
0.1341	3.8335	7000	0.3076	22.3479
0.0895	4.1073	7500	0.3210	22.3593
0.0883	4.3812	8000	0.3223	22.4786
0.0892	4.6550	8500	0.3182	22.1073
0.0937	4.9288	9000	0.3179	21.9008
0.0608	5.2026	9500	0.3326	22.0466
0.0482	5.4765	10000	0.3322	22.0069

Framework versions

• Transformers 4.45.2
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.20.3