whisper-train

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

• Loss: 0.2384
• Wer: 0.2870
• Cer: 0.0749

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 OptimizerNames.RMSPROP and the args are: No additional optimizer arguments
• lr_scheduler_type: linear
• training_steps: 19760
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.1235	1.0006	1000	0.1177	0.3478	0.0882
0.0642	2.0011	2000	0.1340	0.3398	0.0888
0.0337	3.0017	3000	0.1533	0.3365	0.0845
0.0196	4.0022	4000	0.1664	0.3262	0.0841
0.0126	5.0028	5000	0.1725	0.3167	0.0812
0.0085	6.0033	6000	0.1950	0.3327	0.0841
0.0063	7.0039	7000	0.1855	0.3034	0.0807
0.0045	8.0045	8000	0.1932	0.3060	0.0797
0.0033	9.0050	9000	0.1954	0.3015	0.0788
0.0024	10.0056	10000	0.2011	0.3070	0.0800
0.0019	11.0061	11000	0.1976	0.2917	0.0776
0.0015	12.0067	12000	0.2122	0.3028	0.0792
0.0011	13.0072	13000	0.2010	0.2969	0.0779
0.0006	14.0078	14000	0.2187	0.2975	0.0784
0.0005	15.0084	15000	0.2249	0.2956	0.0767
0.0003	16.0089	16000	0.2263	0.2916	0.0761
0.0002	17.0095	17000	0.2314	0.2891	0.0763
0.0001	18.0100	18000	0.2346	0.2886	0.0756
0.0	19.0106	19000	0.2384	0.2870	0.0749

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.1+cu128
• Datasets 4.4.2
• Tokenizers 0.22.2