The model is fine-tuned for ASR in Portuguese. We decided to train in Portuguese because it is a very common language, yet does not have many resources in terms of NLP.

Intended uses & limitations

The model is used for Automatic Speach Recognition. It is fine-tuned in the Portuguese language.

Training and evaluation data

Trained and evaluated on the Common Voice 11 Portuguese data.

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
training_steps: 2000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4076	0.1	200	0.5182	32.4930
0.3462	0.2	400	0.4912	29.0266
0.3283	0.3	600	0.4671	27.0308
0.3579	0.4	800	0.4662	26.6457
0.2766	0.5	1000	0.4639	26.7157
0.2147	1.03	1200	0.4470	26.7857
0.1877	1.13	1400	0.4382	26.4006
0.192	1.23	1600	0.4430	26.3655
0.1894	1.33	1800	0.4349	26.4006
0.1725	1.43	2000	0.4367	26.2605

Framework versions

Transformers 4.26.0.dev0
Pytorch 1.13.0+cu116
Datasets 2.8.1.dev0
Tokenizers 0.13.2

koliskos/whisper-base-pt

Sussurrar

Model description