wav2vec2-large-xls-r-300m-singlish-colab

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

• Loss: 0.7199
• Wer: 0.3337

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: 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: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.2984	4.76	400	2.9046	1.0
1.1895	9.52	800	0.7725	0.4535
0.1331	14.28	1200	0.7068	0.3847
0.0701	19.05	1600	0.7547	0.3617
0.0509	23.8	2000	0.7123	0.3444
0.0385	28.57	2400	0.7199	0.3337

Framework versions

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