kiranpantha/whisper-large-v3-turbo-nepali

Whisper Large v3 Turbo Nepali - Kiran Pantha

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

• Loss: 0.1707
• Wer: 23.6343
• Cer: 5.4903

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: 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: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.3073	0.3597	300	0.2895	53.2870	13.5643
0.2457	0.7194	600	0.2396	45.3704	11.6816
0.166	1.0791	900	0.2062	37.9167	9.6668
0.1477	1.4388	1200	0.1949	37.4306	9.3071
0.1284	1.7986	1500	0.1680	32.6620	8.3235
0.0745	2.1583	1800	0.1706	31.1574	7.5272
0.0701	2.5180	2100	0.1661	32.0370	7.7217
0.0777	2.8777	2400	0.1599	28.6111	7.1308
0.0455	3.2374	2700	0.1723	28.7037	7.0097
0.0375	3.5971	3000	0.1579	26.9444	6.3674
0.0374	3.9568	3300	0.1639	26.8981	6.2794
0.0171	4.3165	3600	0.1711	25.3241	6.2280
0.0219	4.6763	3900	0.1638	25.0	5.9307
0.0089	5.0360	4200	0.1635	24.5139	5.7435
0.0072	5.3957	4500	0.1717	24.1898	5.5711
0.0059	5.7554	4800	0.1707	23.6343	5.4903

Framework versions

• Transformers 4.46.3
• Pytorch 2.5.1+cxx11.abi
• Datasets 3.2.0
• Tokenizers 0.20.3