Hungarians/whisper-small-cv16-hu-v2

Whisper Small Hu v2

This model is a fine-tuned version of openai/whisper-small on the Common Voice 16.0 dataset. It achieves the following results on the evaluation set (at step 12000 best WER&CER modell):

• Loss: 0.1098
• Wer Ortho: 9.3671
• Wer: 8.1026

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: 1.25e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
0.2871	0.33	1000	0.3110	31.4525	28.3160
0.1967	0.67	2000	0.2263	23.7836	20.9280
0.1542	1.0	3000	0.1808	19.5351	16.7032
0.1083	1.34	4000	0.1658	17.2709	14.8236
0.1054	1.67	5000	0.1438	14.9554	12.6653
0.0468	2.01	6000	0.1248	12.8869	10.8034
0.0437	2.34	7000	0.1235	12.2817	10.3913
0.05	2.68	8000	0.1197	11.4958	9.6887
0.022	3.01	9000	0.1119	10.4932	8.9623
0.0286	3.35	10000	0.1141	10.4149	9.0780
0.0233	3.68	11000	0.1150	10.0536	8.6837
0.0124	4.02	12000	0.1098	9.3671	8.1026
0.0165	4.35	13000	0.1143	9.7947	8.3813
0.0174	4.69	14000	0.1136	9.3249	8.0729
0.0107	5.02	15000	0.1150	9.2527	8.2745

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0