wav2vec2-xls-r-common_voice-tr-ft

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

• Loss: 0.3736
• Wer: 0.2930
• Cer: 0.0708

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: 0.0005
• train_batch_size: 12
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 96
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.5462	13.51	500	0.4423	0.4807	0.1188
0.342	27.03	1000	0.3781	0.3954	0.0967
0.2272	40.54	1500	0.3816	0.3595	0.0893
0.1805	54.05	2000	0.3943	0.3487	0.0854
0.1318	67.57	2500	0.3818	0.3262	0.0801
0.1213	81.08	3000	0.3777	0.3113	0.0758
0.0639	94.59	3500	0.3788	0.2953	0.0716

Framework versions

• Transformers 4.14.1
• Pytorch 1.8.0
• Datasets 1.17.0
• Tokenizers 0.10.3