Malaya-speech_fine-tune_realcase_30_Jun_lm

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

• Loss: 0.7669
• Wer: 0.3194

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: 0.0002
• 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
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.2487	1.82	20	0.7188	0.3403
0.6386	3.64	40	0.7061	0.3264
0.3525	5.45	60	0.7403	0.3542
0.3088	7.27	80	0.7483	0.2986
0.2609	9.09	100	0.7669	0.3194

Framework versions

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