meg51/whisper-medium-hindi

Whisper Medium hindi -megha sharma

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

• Loss: 0.3120
• Wer: 18.1765

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 250
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2166	0.8475	250	0.2327	26.1128
0.1217	1.6949	500	0.1955	21.5053
0.0578	2.5424	750	0.2025	20.7536
0.0271	3.3898	1000	0.2230	20.5096
0.0134	4.2373	1250	0.2463	20.3046
0.0105	5.0847	1500	0.2463	19.7970
0.0064	5.9322	1750	0.2636	19.2796
0.0048	6.7797	2000	0.2678	19.5920
0.0034	7.6271	2250	0.2765	19.2991
0.0021	8.4746	2500	0.2710	18.5084
0.0006	9.3220	2750	0.2879	19.2015
0.0001	10.1695	3000	0.2895	18.4303
0.0003	11.0169	3250	0.2930	18.3815
0.0005	11.8644	3500	0.3032	18.5963
0.0001	12.7119	3750	0.3003	18.4889
0.0001	13.5593	4000	0.3054	18.4010
0.0001	14.4068	4250	0.3085	18.2058
0.0	15.2542	4500	0.3104	18.1472
0.0	16.1017	4750	0.3116	18.1863
0.0	16.9492	5000	0.3120	18.1765

Framework versions

• Transformers 4.43.3
• Pytorch 2.4.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1