kiranpantha/whisper-large-v3-nepali

Whisper Large v3 Turbo Nepali - Kiran Pantha

This model is a fine-tuned version of openai/whisper-large-v3 on the OpenSLR54 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0876
• Wer: 18.7250
• Cer: 4.4861

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: 8
• eval_batch_size: 8
• 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: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
0.2266	0.1200	300	11.9034	0.2345	44.7619
0.208	0.2399	600	11.3157	0.2132	41.1060
0.185	0.3599	900	9.4204	0.1753	35.6068
0.1567	0.4798	1200	8.8596	0.1634	33.9324
0.1411	0.5998	1500	8.7004	0.1523	33.0568
0.1377	0.7197	1800	7.3120	0.1371	29.7849
0.1147	0.8397	2100	7.0010	0.1332	27.7112
0.1116	0.9596	2400	6.5798	0.1212	26.3287
0.0757	1.0796	2700	6.1268	0.1193	24.7773
0.0609	1.1995	3000	5.8991	0.1154	24.6237
0.0612	1.3195	3300	5.2599	0.1091	22.0737
0.0627	1.4394	3600	5.3579	0.1045	21.6283
0.0582	1.5594	3900	5.1938	0.0995	21.5054
0.0551	1.6793	4200	4.7947	0.0956	19.8771
0.052	1.7993	4500	4.5473	0.0897	19.1244
0.0438	1.9192	4800	4.4861	0.0876	18.7250

Framework versions

• Transformers 4.47.1
• Pytorch 2.5.1+cxx11.abi
• Datasets 3.2.0
• Tokenizers 0.21.0