jonatasgrosman/whisper-small-pt-cv11-v4

Whisper Small Portuguese

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 pt dataset. It achieves the following results on the evaluation set:

• Loss: 0.3191
• Wer: 14.8844
• Cer: 5.7447

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: 5e-06
• train_batch_size: 32
• eval_batch_size: 16
• 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	Cer
2.9379	0.92	500	0.4783	17.3806	7.0572
2.1727	1.84	1000	0.3721	17.2727	6.7975
1.7856	2.76	1500	0.3466	16.3790	6.4023
1.7803	3.68	2000	0.3372	15.9014	6.2089
1.8312	4.6	2500	0.3303	15.7473	6.0901
1.6403	5.52	3000	0.3256	15.9476	6.1896
1.536	6.45	3500	0.3235	15.5008	6.0928
1.4223	7.37	4000	0.3209	15.3621	6.0735
1.4652	8.29	4500	0.3209	15.2696	5.9326
1.2572	9.21	5000	0.3191	14.8844	5.7447
1.7142	10.13	5500	0.3182	15.0077	5.8469
1.4195	11.05	6000	0.3171	15.0693	5.8856
1.3965	11.97	6500	0.3167	15.0539	5.8580

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1+cu116
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2