Sercan/whisper-small-tr

Whisper Small Turkish

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 tr dataset. It achieves the following results on the evaluation set:

• Loss: 0.2799
• Wer: 17.2753
• Cer: 4.5335

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: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.1044	1.07	1000	0.2777	18.4046	4.8810
0.0469	3.02	2000	0.2799	17.2753	4.5335
0.014	4.09	3000	0.3202	18.0800	4.9039
0.0039	6.04	4000	0.3326	18.2964	5.0192
0.0022	7.11	5000	0.3453	18.0307	4.9470

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.8.1.dev0
• Tokenizers 0.13.2