arun100/whisper-base-hi-3

Whisper Base Hindi

This model is a fine-tuned version of arun100/whisper-base-hi-2 on the google/fleurs hi_in dataset. It achieves the following results on the evaluation set:

• Loss: 0.4468
• Wer: 27.7206

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: 5e-07
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• 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
0.4805	33.0	250	0.4868	30.4186
0.3559	66.0	500	0.4417	29.0909
0.2655	99.0	750	0.4307	28.2165
0.1987	133.0	1000	0.4350	27.8326
0.1472	166.0	1250	0.4468	27.7206
0.1061	199.0	1500	0.4640	28.0992
0.0767	233.0	1750	0.4835	28.5737
0.0541	266.0	2000	0.5032	28.6857
0.0396	299.0	2250	0.5202	28.7763
0.03	333.0	2500	0.5353	29.2029
0.0237	366.0	2750	0.5479	28.9096
0.0195	399.0	3000	0.5587	28.9096
0.0163	433.0	3250	0.5683	28.9469
0.014	466.0	3500	0.5767	29.1336
0.0121	499.0	3750	0.5838	29.3415
0.0108	533.0	4000	0.5900	29.2775
0.01	566.0	4250	0.5951	29.6081
0.0093	599.0	4500	0.5988	29.4855
0.0088	633.0	4750	0.6012	29.5281
0.0087	666.0	5000	0.6020	29.4268

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.16.2.dev0
• Tokenizers 0.15.0