AndrewMcDowell/wav2vec2-xls-r-1b-arabic

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - AR dataset. It achieves the following results on the evaluation set:

• Loss: 1.1373
• Wer: 0.8607

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: 6.5e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 30.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.2416	0.84	500	1.2867	0.8875
2.3089	1.67	1000	1.8336	0.9548
2.3614	2.51	1500	1.5937	0.9469
2.5234	3.35	2000	1.9765	0.9867
2.5373	4.19	2500	1.9062	0.9916
2.5703	5.03	3000	1.9772	0.9915
2.4656	5.86	3500	1.8083	0.9829
2.4339	6.7	4000	1.7548	0.9752
2.344	7.54	4500	1.6146	0.9638
2.2677	8.38	5000	1.5105	0.9499
2.2074	9.21	5500	1.4191	0.9357
2.3768	10.05	6000	1.6663	0.9665
2.3804	10.89	6500	1.6571	0.9720
2.3237	11.72	7000	1.6049	0.9637
2.317	12.56	7500	1.5875	0.9655
2.2988	13.4	8000	1.5357	0.9603
2.2906	14.24	8500	1.5637	0.9592
2.2848	15.08	9000	1.5326	0.9537
2.2381	15.91	9500	1.5631	0.9508
2.2072	16.75	10000	1.4565	0.9395
2.197	17.59	10500	1.4304	0.9406
2.198	18.43	11000	1.4230	0.9382
2.1668	19.26	11500	1.3998	0.9315
2.1498	20.1	12000	1.3920	0.9258
2.1244	20.94	12500	1.3584	0.9153
2.0953	21.78	13000	1.3274	0.9054
2.0762	22.61	13500	1.2933	0.9073
2.0587	23.45	14000	1.2516	0.8944
2.0363	24.29	14500	1.2214	0.8902
2.0302	25.13	15000	1.2087	0.8871
2.0071	25.96	15500	1.1953	0.8786
1.9882	26.8	16000	1.1738	0.8712
1.9772	27.64	16500	1.1647	0.8672
1.9585	28.48	17000	1.1459	0.8635
1.944	29.31	17500	1.1414	0.8616

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0