wav2vec2-xls-r-timit-trainer

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: 1.1064
• 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: 0.0003
• 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: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.5537	4.03	500	0.6078	1.0
0.5444	8.06	1000	0.4990	0.9994
0.3744	12.1	1500	0.5530	1.0
0.2863	16.13	2000	0.6401	1.0
0.2357	20.16	2500	0.6485	1.0
0.1933	24.19	3000	0.7448	0.9994
0.162	28.22	3500	0.7502	1.0
0.1325	32.26	4000	0.7801	1.0
0.1169	36.29	4500	0.8334	1.0
0.1031	40.32	5000	0.8269	1.0
0.0913	44.35	5500	0.8432	1.0
0.0793	48.39	6000	0.8738	1.0
0.0694	52.42	6500	0.8897	1.0
0.0613	56.45	7000	0.8966	1.0
0.0548	60.48	7500	0.9398	1.0
0.0444	64.51	8000	0.9548	1.0
0.0386	68.55	8500	0.9647	1.0
0.0359	72.58	9000	0.9901	1.0
0.0299	76.61	9500	1.0151	1.0
0.0259	80.64	10000	1.0526	1.0
0.022	84.67	10500	1.0754	1.0
0.0189	88.71	11000	1.0688	1.0
0.0161	92.74	11500	1.0914	1.0
0.0138	96.77	12000	1.1064	1.0

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.13.3
• Tokenizers 0.10.3