sparshrestha/finetuned-whisper-small-nepali

Fine-tuned Whisper Small Nepali

This model is a fine-tuned version of openai/whisper-small on the Common Voice 26.0 + Google FLEURS ne_np dataset. It achieves the following results on the evaluation set:

• Loss: 0.4616
• Wer: 63.1340
• Normalized Wer: 62.4060
• Cer: 27.8890
• Avg Pred Words: 10.0194
• Avg Ref Words: 11.3694
• Empty Prediction Rate: 0.0

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 50
• training_steps: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Normalized Wer	Cer	Avg Pred Words	Avg Ref Words	Empty Prediction Rate
13.3546	0.8686	100	0.7885	86.2872	85.6150	39.2049	9.7677	11.3694	0.0
5.8797	1.7296	200	0.5593	70.1075	69.7120	31.0429	9.8871	11.3694	0.0
4.4283	2.5907	300	0.4871	65.2426	64.6191	28.6433	10.0581	11.3694	0.0
3.8697	3.4517	400	0.4643	63.5198	62.8316	28.1000	9.9887	11.3694	0.0
3.4989	4.3127	500	0.4616	63.1340	62.4060	27.8890	10.0194	11.3694	0.0

Framework versions

• Transformers 5.12.1
• Pytorch 2.11.0+cu128
• Datasets 5.0.0
• Tokenizers 0.22.2