wav2vec2-large-xls-r-300m-amharic-ASR-final

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_16_1 dataset. It achieves the following results on the evaluation set:

• Loss: 1.8430
• Wer: 0.9093
• Cer: 0.3582

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: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 80
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
12.9379	5.0	100	4.1617	1.0	0.9997
4.0909	10.0	200	4.0177	1.0	0.9531
3.9876	15.0	300	3.9620	1.0	0.9404
3.9269	20.0	400	3.9533	0.9987	0.9172
3.095	25.0	500	2.1131	1.0173	0.5117
0.9792	30.0	600	1.7629	0.9373	0.4255
0.4635	35.0	700	1.8393	0.9560	0.4060
0.3034	40.0	800	1.8041	0.9099	0.3959
0.2285	45.0	900	1.8211	0.9173	0.3769
0.189	50.0	1000	1.8982	0.9306	0.3854
0.1692	55.0	1100	1.8831	0.9393	0.3795
0.1511	60.0	1200	1.8967	0.9346	0.3739
0.1326	65.0	1300	1.9037	0.9420	0.3823
0.1185	70.0	1400	1.8853	0.9146	0.3653
0.1102	75.0	1500	1.8515	0.9079	0.3598
0.1059	80.0	1600	1.8430	0.9093	0.3582

Framework versions

• Transformers 4.42.0
• Pytorch 2.3.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1