wav2vec2-xls-r-300m-th-beta

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.9096
• Wer: 0.7261
• Cer: 0.2160
• Clean Cer: 0.1909

Model description

This model is available until Jan 12, 2023

Intended uses & limitations

More information needed

Training and evaluation data

We use our custom dataset splited into 70k training dataset, and 7k evaluation dataset This is our detailed dataset

1. Common Voice 11

• filter 5k fifties age males out
• remain 25k training dataset

2. Botnoi voice

• 45k training dataset

Both dataset was through our custom cleansing text data.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• eval_batch_size: 16
• 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
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Clean Cer
6.9481	0.34	500	3.5952	1.0	0.9815	0.9779
2.0387	0.68	1000	0.9096	0.7261	0.2160	0.1909

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu116
• Datasets 2.8.0
• Tokenizers 0.13.2