satarupa22/Wishper-small-asr-bn

Whisper Small EN-BN ASR

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

• Loss: 0.0514
• Wer: 17.3903

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 16
• 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_ratio: 0.03
• training_steps: 8000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1073	0.3811	1000	0.0995	31.1461
0.0831	0.7622	2000	0.0781	26.3300
0.0525	1.1433	3000	0.0689	24.3359
0.0435	1.5244	4000	0.0605	21.2507
0.0391	1.9055	5000	0.0543	19.3544
0.018	2.2866	6000	0.0564	18.5793
0.0175	2.6677	7000	0.0519	18.1804
0.0063	3.0488	8000	0.0514	17.3903

Framework versions

• Transformers 4.50.3
• Pytorch 2.4.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1