model-1h

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

• Loss: 1.8317
• 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: 5
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
11.4106	1.24	10	7.1597	1.0
4.777	2.47	20	3.9782	1.0
3.6585	3.71	30	3.3961	1.0
3.3143	4.94	40	3.1481	1.0
3.3318	6.24	50	3.0596	1.0
3.1368	7.47	60	2.9751	1.0
3.1058	8.71	70	2.9510	1.0
3.0605	9.94	80	2.9479	1.0
3.2043	11.24	90	2.9270	1.0
3.0424	12.47	100	2.9349	1.0
3.0374	13.71	110	2.9316	1.0
3.0256	14.94	120	2.9165	1.0
3.1724	16.24	130	2.9076	1.0
3.0119	17.47	140	2.9034	1.0
2.9937	18.71	150	2.8812	1.0
2.9775	19.94	160	2.8674	1.0
3.0826	21.24	170	2.8147	1.0
2.8717	22.47	180	2.7212	1.0
2.7714	23.71	190	2.6149	0.9952
2.634	24.94	200	2.4611	0.9984
2.5637	26.24	210	2.2734	1.0
2.237	27.47	220	2.0705	1.0
2.0381	28.71	230	1.9216	1.0
1.8788	29.94	240	1.8317	1.0

Framework versions

• Transformers 4.11.3
• Pytorch 1.9.1+cu102
• Datasets 1.18.3
• Tokenizers 0.10.3