wav2vec2-xls-r-300m_Mrbrown_finetune1

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the uob_singlish dataset.

This time use self-made dataset(cut the audio of "https://www.youtube.com/watch?v=a2ZOTD3R7JI" into slices and write the corresponding transcript, totally 4 mins), don't know why the word-error-rate keep 1. But can know that much be the problem of dataset, because last time use the same pre-trained model and standard singlish corpus fine-tune get nice result. (can find it at:RuiqianLi/wav2vec2-large-xls-r-300m-singlish-colab)

It achieves the following results on the evaluation set:

• Loss: 3.0927
• Wer: 1.0

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: 0.01
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.7943	20.0	200	3.0597	1.0
2.9902	40.0	400	3.1604	1.0
2.9696	60.0	600	3.1112	1.0
2.8885	80.0	800	3.0234	1.0
2.8154	100.0	1000	3.0927	1.0

Framework versions

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