Bidwill/whisper-medium-sanskrit-try-2

Whisper medium sanskrit try - Bidit Sadhukhan

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

• Loss: 0.1326
• Wer: 24.8767

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: 6.25e-06
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: reduce_lr_on_plateau
• lr_scheduler_warmup_steps: 500
• training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0708	0.23	500	0.1239	35.9144
0.0516	0.47	1000	0.1093	31.4377
0.0492	0.7	1500	0.1031	28.9836
0.0472	0.93	2000	0.1003	27.9471
0.027	1.17	2500	0.1078	27.6726
0.0246	1.4	3000	0.0959	25.6948
0.0286	1.63	3500	0.1000	25.3138
0.0235	1.86	4000	0.0980	25.1513
0.0113	2.1	4500	0.1035	24.4453
0.017	2.33	5000	0.1038	25.0896
0.0171	2.56	5500	0.1038	25.1121
0.017	2.8	6000	0.1105	25.5603
0.0065	3.03	6500	0.1182	25.4370
0.0078	3.26	7000	0.1247	25.2409
0.0111	3.5	7500	0.1304	26.5464
0.0102	3.73	8000	0.1191	25.8909
0.0155	3.96	8500	0.1142	25.2073
0.006	4.2	9000	0.1269	24.9496
0.0074	4.43	9500	0.1335	25.1513
0.0053	4.66	10000	0.1326	24.8767

Framework versions

• Transformers 4.34.0.dev0
• Pytorch 2.0.1+cu117
• Datasets 2.14.5
• Tokenizers 0.14.0