malaya-speech_Mrbrown_finetune1

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

malaya-speech_Mrbrown_finetune1

This model is a fine-tuned version of malay-huggingface/wav2vec2-xls-r-300m-mixed 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), get really bad fine-tuning result, that may mean the training/fine-tuning dataset must be high quality/at least several hours? Or maybe is because the learning rate is set too high(0.01) ? Still searching for the important factors.

It achieves the following results on the evaluation set:

• Loss: 3.8458
• Wer: 1.01

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

Framework versions

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