ALM/whisper-hu-small-augmented

Whisper Small Hungarian - Robust

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

• Loss: 0.5243
• Wer: 30.9042

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: 128
• eval_batch_size: 64
• 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
0.0763	7.46	500	0.5268	40.5099
0.0147	14.93	1000	0.5233	36.1429
0.0064	22.39	1500	0.5467	35.0934
0.0045	29.85	2000	0.5434	34.2929
0.0019	37.31	2500	0.5348	32.7868
0.0008	44.78	3000	0.5314	32.0605
0.0008	52.24	3500	0.5438	32.6920
0.0005	59.7	4000	0.5428	32.0931
0.0003	67.16	4500	0.5328	31.2511
0.0004	74.63	5000	0.5292	31.1236

Framework versions

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