ixxan/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 test set of THUGY20:

• Loss: 0.7473
• Wer Ortho: 18.0908
• Wer: 17.9401
• Cer: 4.9274

Training procedure

Finetuning code avaiblable in https://github.com/ixxan/ug-speech

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• 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: 50
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer	Cer
0.3815	0.8058	500	0.7944	34.8819	34.7960	10.4265
0.1343	1.6116	1000	0.7441	28.3393	28.3550	8.3051
0.0646	2.4174	1500	0.7396	27.7378	27.5653	8.5366
0.0311	3.2232	2000	0.6984	25.1910	24.9445	7.5643
0.0176	4.0290	2500	0.6934	21.3709	21.2523	5.8316
0.0075	4.8348	3000	0.7654	20.5541	20.3603	5.7519
0.0023	5.6406	3500	0.7686	18.7582	18.5846	5.1923
0.0004	6.4464	4000	0.7473	18.0908	17.9401	4.9274

Framework versions

• Transformers 4.46.2
• Pytorch 2.5.1+cu121
• Datasets 3.1.0
• Tokenizers 0.20.3