ali-issa/FYP_ARABIZI

wav2vec2-Arabizi-gpu-colab-similar-to-german-param

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5609
• Wer: 0.4042

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: 0.0001
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 6
• total_train_batch_size: 12
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.6416	2.83	400	2.8983	1.0
1.4951	5.67	800	0.6272	0.6097
0.6419	8.51	1200	0.5491	0.5069
0.4767	11.35	1600	0.5152	0.4553
0.3899	14.18	2000	0.5436	0.4475
0.3342	17.02	2400	0.5400	0.4431
0.2982	19.85	2800	0.5599	0.4248
0.2738	22.69	3200	0.5401	0.4103
0.2563	25.53	3600	0.5710	0.4198
0.2443	28.37	4000	0.5609	0.4042

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu113
• Datasets 1.18.3
• Tokenizers 0.10.3