Mehtap/base_03

base Turkish Whisper (bTW)

This model is a fine-tuned version of openai/whisper-base on the Ermetal Meetings dataset. It achieves the following results on the evaluation set:

• Loss: 2.0552
• Wer: 1.3802
• Cer: 0.8297

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: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.3477	33.33	100	1.8981	1.2433	0.8110
0.0238	66.67	200	1.7919	0.9340	0.5818
0.0032	100.0	300	1.8780	0.9756	0.6155
0.0014	133.33	400	1.9332	1.3582	0.8039
0.0008	166.67	500	1.9769	1.6333	0.9329
0.0005	200.0	600	2.0099	1.3790	0.8230
0.0004	233.33	700	2.0307	1.3851	0.8270
0.0004	266.67	800	2.0442	1.3851	0.8286
0.0003	300.0	900	2.0523	1.3814	0.8303
0.0003	333.33	1000	2.0552	1.3802	0.8297

Framework versions

• Transformers 4.26.0
• Pytorch 1.12.0+cu102
• Datasets 2.9.0
• Tokenizers 0.13.2