sevda-tatlih/whisper-small-uyghur-thugy20

Whisper Small Fine-tuned with THUYG20 Uyghur Dataset

This model is a fine-tuned version of openai/whisper-small on the THUGY20: A free Uyghur speech database dataset. It achieves the following results on the evaluation set:

• Loss: 0.3708
• Wer Ortho: 22.9277
• Wer: 22.8471
• Cer: 5.8681

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: 3e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 200
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer	Wer Ortho
0.5311	0.8403	200	9.8849	0.5551	40.8203	40.8026
0.2102	1.6807	400	7.9473	0.4169	31.2599	31.2051
0.0712	2.5210	600	7.7075	0.3970	28.7094	28.7809
0.0227	3.3613	800	7.1401	0.3966	26.4656	26.4852
0.0109	4.2017	1000	6.7159	0.3661	24.5683	24.6218
0.0067	5.0420	1200	6.1440	0.3753	23.9434	24.0318
0.003	5.8824	1400	5.9610	0.3694	23.1822	23.2315
0.002	6.7227	1600	5.8850	0.3728	22.8925	22.9686
0.0017	7.5630	1800	5.8695	0.3708	22.8584	22.9394
0.0018	8.4034	2000	5.8681	0.3710	22.8471	22.9277

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0