HowMannyMore/whisper-small-urdu

whisper-small-ur

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

• Loss: 0.4843
• Wer: 33.3110

Training and evaluation data

Dataset included two rows; transcription & audio. The model was prepared using a dataset of 6500 rows. Train-test split was applied, 82% training (5324) and 18% testing (1176).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• 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: 1200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5907	0.6	400	0.6646	44.5644
0.2862	1.2	800	0.5806	38.1544
0.251	1.8	1200	0.4843	33.3110

Framework versions

• Transformers 4.36.0.dev0
• Pytorch 2.1.0+cu118
• Datasets 2.14.6
• Tokenizers 0.14.1