wav2vec2-large-xls-r-300m-tr-colab

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: 0.4316
• Wer: 0.2905

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
3.9953	3.67	400	0.7024	0.7226
0.4046	7.34	800	0.4342	0.5343
0.201	11.01	1200	0.4396	0.5290
0.1513	14.68	1600	0.4319	0.4108
0.1285	18.35	2000	0.4422	0.3864
0.1086	22.02	2400	0.4568	0.3796
0.0998	25.69	2800	0.4687	0.3732
0.0863	29.36	3200	0.4726	0.3803
0.0809	33.03	3600	0.4479	0.3601
0.0747	36.7	4000	0.4624	0.3525
0.0692	40.37	4400	0.4366	0.3435
0.0595	44.04	4800	0.4204	0.3510
0.0584	47.71	5200	0.4202	0.3402
0.0545	51.38	5600	0.4366	0.3343
0.0486	55.05	6000	0.4492	0.3678
0.0444	58.72	6400	0.4471	0.3301
0.0406	62.39	6800	0.4382	0.3318
0.0341	66.06	7200	0.4295	0.3258
0.0297	69.72	7600	0.4336	0.3205
0.0295	73.39	8000	0.4240	0.3199
0.0261	77.06	8400	0.4316	0.3143
0.0247	80.73	8800	0.4300	0.3165
0.0207	84.4	9200	0.4380	0.3111
0.0203	88.07	9600	0.4218	0.2998
0.0174	91.74	10000	0.4271	0.2973
0.015	95.41	10400	0.4330	0.2939
0.0144	99.08	10800	0.4316	0.2905

Framework versions

• Transformers 4.23.1
• Pytorch 1.12.1+cu102
• Datasets 2.5.2
• Tokenizers 0.13.1