whisper-base-me

This model is a fine-tuned version of openai/whisper-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5421
• Mazanderani Wer: 0.6836
• Mazanderani Cer: 0.2510
• Gilaki Wer: 0.9606
• Gilaki Cer: 0.3818
• Zazaki Wer: 0.7623
• Zazaki Cer: 0.3182
• Laki Kurdish Wer: 0.6037
• Laki Kurdish Cer: 0.1653
• Talysh Wer: 0.9167
• Talysh Cer: 0.5
• Hawrami Wer: 0.4145
• Hawrami Cer: 0.0865
• Southern Kurdish Wer: 0.5610
• Southern Kurdish Cer: 0.1721
• Avg Wer: 0.7003
• Avg Cer: 0.2678

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: 128
• eval_batch_size: 128
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 4.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Mazanderani Wer	Mazanderani Cer	Gilaki Wer	Gilaki Cer	Zazaki Wer	Zazaki Cer	Laki Kurdish Wer	Laki Kurdish Cer	Talysh Wer	Talysh Cer	Hawrami Wer	Hawrami Cer	Southern Kurdish Wer	Southern Kurdish Cer	Avg Wer	Avg Cer
0.5933	1.0	164	0.7337	0.8884	0.3139	0.9969	0.3853	0.9485	0.7335	0.7857	0.2461	0.9167	0.5	0.5257	0.1162	0.6900	0.2137	0.8217	0.3584
0.4436	2.0	328	0.5930	0.7514	0.2939	0.9723	0.3818	0.8382	0.4153	0.6487	0.1846	0.9167	0.5	0.4419	0.0933	0.5951	0.1798	0.7378	0.2927
0.387	3.0	492	0.5509	0.6821	0.2436	0.9523	0.3735	0.7892	0.3340	0.6094	0.1678	0.9167	0.5	0.4230	0.0891	0.5705	0.1871	0.7062	0.2707
0.3575	4.0	656	0.5421	0.6836	0.2510	0.9606	0.3818	0.7623	0.3182	0.6037	0.1653	0.9167	0.5	0.4145	0.0865	0.5610	0.1721	0.7003	0.2678

Framework versions

• Transformers 4.49.0.dev0
• Pytorch 2.6.0+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0