wav2vec2-large-xlsr-53-german-cv13-restart

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the MOZILLA-FOUNDATION/COMMON_VOICE_13_0 - DE dataset. It achieves the following results on the evaluation set:

• Loss: 0.1121
• Wer: 0.1075
• Cer: 0.0286

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: 0.0003
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.01
• num_epochs: 15.0

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
0.1679	1.0	4348	0.0459	0.1617	0.1707
0.1597	2.0	8697	0.0479	0.1592	0.1717
0.1364	3.0	13045	0.0425	0.1524	0.1563
0.1311	4.0	17394	0.0406	0.1446	0.1515
0.1152	5.0	21742	0.0397	0.1431	0.1470
0.107	6.0	26091	0.0369	0.1382	0.1377
0.0957	7.0	30439	0.0373	0.1343	0.1372
0.0924	8.0	34788	0.0355	0.1335	0.1315
0.0835	9.0	39136	0.0384	0.1328	0.1380
0.0775	10.0	43485	0.0328	0.1232	0.1229
0.0752	11.0	47833	0.0309	0.1220	0.1174
0.0691	12.0	52182	0.0327	0.1182	0.1197
0.0689	13.0	56530	0.0307	0.1163	0.1150
0.0653	14.0	60879	0.0304	0.1141	0.1126
0.0717	15.0	65220	0.1121	0.1075	0.0286

Framework versions

• Transformers 4.34.0
• Pytorch 2.1.0+cu121
• Datasets 2.14.5
• Tokenizers 0.14.0