whisper-small-ug

This model is a fine-tuned version of openai/whisper-small. The model is trained on transcripts written in Uyghur Latin Script via utilising Uzbek Tokeniser , as Uyghur Tokeniser is not included in Whisper. Therefore, the output of the model is in Uyghur Latin Script. To convert the output to the Uyghur Arabic Script, you can use the Uyghur script converter: https://github.com/neouyghur/ScriptConverter4Uyghur

or you can use online script converter: https://www.yulghun.com/imla/convert.html

It achieves the following results on the evaluation set:

• Loss: 0.3563
• Wer: 26.8793

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: 8
• 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: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2677	1.43	1000	0.4063	34.1157
0.1035	2.85	2000	0.3375	29.2183
0.0226	4.28	3000	0.3472	27.5155
0.0073	5.71	4000	0.3563	26.8793

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.1+cu121
• Datasets 2.15.0
• Tokenizers 0.15.0