ALM/whisper-el-small-augmented

Whisper Small Greek - Robust

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

• Loss: 0.3605
• Wer: 24.5264

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: 16
• eval_batch_size: 8
• 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2223	2.35	500	0.4403	37.9922
0.0908	4.69	1000	0.4041	35.6519
0.0465	7.04	1500	0.4189	34.3053
0.0168	9.39	2000	0.3972	29.9127
0.0118	11.74	2500	0.4043	28.9933
0.0053	14.08	3000	0.3968	28.5940
0.0032	16.43	3500	0.3664	25.6779
0.0009	18.78	4000	0.3665	26.2444
0.0003	21.13	4500	0.3620	25.2879
0.0004	23.47	5000	0.3570	24.8607

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.7.1
• Tokenizers 0.13.2