Whisper Large v2 PL

This model is a fine-tuned version of bardsai/whisper-large-v2-pl on the Common Voice 11.0 and the FLEURS datasets. It achieves the following results on the evaluation set:

Loss: 0.3684
Wer: 7.2802

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 8
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
training_steps: 2100
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0047	1.35	700	0.3428	8.5562
0.0011	2.7	1400	0.3605	7.5505
0.0003	4.05	2100	0.3684	7.2802

Framework versions

Transformers 4.26.0.dev0
Pytorch 1.13.0+cu117
Datasets 2.7.1.dev0
Tokenizers 0.13.2

Downloads last month: 77

Automatic Speech Recognition

This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Datasets used to train bardsai/whisper-large-v2-pl-v2

Collection including bardsai/whisper-large-v2-pl-v2

Polish Whisper

Collection

Whisper ASR models fine-tuned for Polish Language • 5 items • Updated Jan 10

Evaluation results

WER on Common Voice 11.0
test set self-reported

7.280
WER on Common Voice 11.0
test set self-reported

7.310
WER unnormalized on Common Voice 11.0
test set self-reported

20.180
CER on Common Voice 11.0
test set self-reported

2.080
MER on Common Voice 11.0
test set self-reported

7.270
WER on facebook/voxpopuli
test set self-reported

9.610
WER unnormalized on facebook/voxpopuli
test set self-reported

30.330
CER on facebook/voxpopuli
test set self-reported

5.500
MER on facebook/voxpopuli
test set self-reported

9.450
WER on google/fleurs
test set self-reported

8.680

View on Papers With Code