whisper-tiny-final

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

• Loss: 0.0714
• Wer: 6.3947

Model description

Step Training Loss Validation Loss Wer 1000 0.727300 0.734777 71.347666 2000 0.392000 0.430395 52.059163 3000 0.317100 0.305939 39.781162 4000 0.206400 0.225029 30.785726 5000 0.152800 0.169434 23.076923 6000 0.119000 0.130408 16.517293 7000 0.082300 0.102279 11.755650 8000 0.079600 0.085155 8.511574 9000 0.051400 0.075068 7.048991 10000 0.045000 0.071429 6.394678

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.7273	1.6051	1000	0.7348	71.3477
0.392	3.2103	2000	0.4304	52.0592
0.3171	4.8154	3000	0.3059	39.7812
0.2064	6.4205	4000	0.2250	30.7857
0.1528	8.0257	5000	0.1694	23.0769
0.119	9.6308	6000	0.1304	16.5173
0.0823	11.2360	7000	0.1023	11.7556
0.0796	12.8411	8000	0.0852	8.5116
0.0514	14.4462	9000	0.0751	7.0490
0.045	16.0514	10000	0.0714	6.3947

Framework versions

• Transformers 4.40.2
• Pytorch 2.2.0
• Datasets 2.19.1
• Tokenizers 0.19.1