whisper-medium-pa

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

• eval_loss: 0.0371
• eval_wer: 12.9712
• eval_runtime: 4735.7351
• eval_samples_per_second: 1.271
• eval_steps_per_second: 0.159
• epoch: 0.2612
• step: 15000

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: 15000
• mixed_precision_training: Native AMP

Initial Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0823	1.2804	1000	0.1072	28.8516
0.0501	2.5608	2000	0.0952	25.4005
0.0287	3.8412	3000	0.1008	24.5475
0.0088	5.1216	4000	0.1249	24.1288
0.0046	6.4020	5000	0.1389	23.5358
0.0023	7.6825	6000	0.1535	23.5254
0.0011	8.9629	7000	0.1537	23.3668
0.0001	10.2433	8000	0.1720	23.0677
0.0002	11.5237	9000	0.1789	22.7452
0.0	12.8041	10000	0.1804	22.7790

Re-Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.067700	-	1000	0.1072	14.635525
0.055900	-	2000	0.0952	14.542690
0.050000	-	3000	0.1008	14.295132
0.043000	-	4000	0.1249	13.520168
0.043100	-	5000	0.1389	12.971235

Framework versions

• Transformers 4.47.1
• Pytorch 2.1.0+cu118
• Datasets 3.2.0
• Tokenizers 0.21.0