alkiskoudounas/whisper-el-medium-augmented-1

Whisper Medium Greek - Robust

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

• Loss: 0.4312
• Wer: 24.3221

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-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0905	2.35	1000	0.5419	37.8343
0.0552	4.69	2000	0.5118	34.6954
0.0329	7.04	3000	0.5332	32.3180
0.0219	9.39	4000	0.5185	30.1913
0.0161	11.74	5000	0.4908	32.3366
0.0063	14.08	6000	0.4741	28.4733
0.0015	16.43	7000	0.4400	26.3187
0.001	18.78	8000	0.4428	25.5293
0.0001	21.13	9000	0.4382	25.2043
0.0014	23.47	10000	0.4328	24.3221

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.11.0+cu113
• Datasets 2.7.1
• Tokenizers 0.12.1