wav2vec2-large-xlsr-53-demo-colab

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.6901
• Wer: 1.6299

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: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.5034	3.42	400	3.5852	1.0
1.7853	6.83	800	0.7430	1.6774
0.5675	10.26	1200	0.6513	1.6330
0.3761	13.67	1600	0.6208	1.6081
0.2776	17.09	2000	0.6401	1.6081
0.2266	20.51	2400	0.6410	1.6295
0.1949	23.93	2800	0.6910	1.6287
0.1672	27.35	3200	0.6901	1.6299

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.14.0
• Tokenizers 0.10.3