Sercan/whisper-small-tr-2

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: 0.2576
• Wer: 16.6327
• Cer: 4.2853

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
0.1412	0.08	400	0.2656	19.8335	5.2393
0.0851	1.03	800	0.2382	18.6300	4.8916
0.0525	1.11	1200	0.2532	19.1696	5.2238
0.0163	2.07	1600	0.2447	17.2014	4.5840
0.0202	3.02	2000	0.2472	17.1063	4.4935
0.0075	3.1	2400	0.2503	17.0151	4.4318
0.0039	4.05	2800	0.2514	16.7433	4.3655
0.0038	5.01	3200	0.2565	16.8870	4.3582
0.0023	5.09	3600	0.2590	16.6987	4.3337
0.0013	6.04	4000	0.2576	16.6327	4.2853
0.0011	6.12	4400	0.2647	16.9122	4.3556
0.001	7.07	4800	0.2615	16.6346	4.2839

Framework versions

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