ihanif/whisper-turbo-ar

Whisper Small PS - Hanif Rahman

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5707
• Wer Ortho: 40.7188
• Wer: 40.0571

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: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_steps: 200
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
0.9719	0.2268	100	0.8098	59.2165	59.0924
0.8427	0.4535	200	0.7384	55.1748	54.5596
0.7493	0.6803	300	0.6743	48.8614	48.3473
0.684	0.9070	400	0.6384	46.1094	45.5534
0.4819	1.1338	500	0.6348	44.3341	43.7123
0.4777	1.3605	600	0.6026	43.6758	42.9264
0.4433	1.5873	700	0.5789	41.7386	40.9991
0.446	1.8141	800	0.5647	40.2709	39.5995
0.3166	2.0408	900	0.5681	40.4490	39.7771
0.3187	2.2676	1000	0.5707	40.7188	40.0571

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.0