wav2vec2-large-xls-r-300m-amharic-demo-colab

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: 1.6333
• Wer: 0.8639

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: 60
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
12.6948	5.0	100	4.1621	1.0
4.1026	10.0	200	4.0365	1.0
4.0037	15.0	300	3.9726	1.0007
3.9485	20.0	400	3.9524	1.0007
3.4635	25.0	500	2.4384	0.9980
1.1709	30.0	600	1.6987	0.9453
0.4955	35.0	700	1.5927	0.9073
0.3163	40.0	800	1.6750	0.8833
0.2372	45.0	900	1.6683	0.8813
0.1896	50.0	1000	1.6555	0.8779
0.1619	55.0	1100	1.6312	0.8819
0.1473	60.0	1200	1.6333	0.8639

Framework versions

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