Whisper Large Swedish

This model is a fine-tuned version of openai/whisper-large-v2 trained on NST Swedish ASR and evaluated on Common Voice 11 testset. It achieves the following results on the evaluation set

Loss: 0.2337
Wer: 9.2206

Model description

openai/whisper-large-v2 had a WER of 10.6 on Common Voice 9 testset.

Intended uses & limitations

More information needed

Training and evaluation data

The training dataset contains 276 000 examples and with a batch size of 64 and training 5000 it is 1.14 epochs. More training data or more epochs would probably improve the result even further.

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: 5000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0695	0.2	1000	0.2695	12.4671
0.0524	0.4	2000	0.2659	11.6367
0.046	0.6	3000	0.2402	10.6557
0.0342	0.8	4000	0.2339	10.1774
0.0224	1.14	5000	0.2337	9.2206

Framework versions

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

bjelkenhed
/

whisper-large-sv