jonatasgrosman/whisper-small-pt-cv11-v4_2

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: 0.3023
• Wer: 14.2373
• Cer: 5.5236

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
1.1113	0.92	500	0.3897	16.8721	6.7919
0.9009	1.84	1000	0.3318	15.9322	6.2310
0.7631	2.76	1500	0.3177	15.4854	5.8939
0.7163	3.68	2000	0.3130	14.8998	5.7972
0.6334	4.6	2500	0.3034	14.7920	5.6867
0.5746	5.52	3000	0.3029	14.6225	5.6397
0.5359	6.45	3500	0.3018	14.4838	5.5789
0.5058	7.37	4000	0.3010	14.5917	5.6839
0.4833	8.29	4500	0.3023	14.2373	5.5236
0.4398	9.21	5000	0.3005	14.4222	5.5844
0.4359	10.13	5500	0.2999	14.4838	5.6259
0.4036	11.05	6000	0.2995	14.2835	5.5623

Framework versions

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