wav2vec2-xls-r-common_voice-tr-ft-500sh

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: 0.5794
• Wer: 0.4009
• Cer: 0.1032

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.5288	17.0	500	0.5099	0.5426	0.1432
0.2967	34.0	1000	0.5421	0.4746	0.1256
0.2447	51.0	1500	0.5347	0.4831	0.1267
0.122	68.01	2000	0.5854	0.4479	0.1161
0.1035	86.0	2500	0.5597	0.4457	0.1166
0.081	103.0	3000	0.5748	0.4250	0.1144
0.0849	120.0	3500	0.5598	0.4337	0.1145
0.0542	137.01	4000	0.5687	0.4223	0.1097
0.0318	155.0	4500	0.5904	0.4057	0.1052
0.0106	172.0	5000	0.5794	0.4009	0.1032

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2
• Datasets 1.18.2
• Tokenizers 0.10.3