wav2vec2-large-xls-hun-53h-colab

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.6027
• Wer: 0.4618

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: 23
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
13.4225	0.67	100	3.7750	1.0
3.4121	1.34	200	3.3166	1.0
3.2263	2.01	300	3.1403	1.0
3.0038	2.68	400	2.2474	0.9990
1.2243	3.35	500	0.8174	0.7666
0.6368	4.03	600	0.6306	0.6633
0.4426	4.7	700	0.6151	0.6648
0.3821	5.37	800	0.5765	0.6138
0.3337	6.04	900	0.5522	0.5785
0.2832	6.71	1000	0.5822	0.5691
0.2485	7.38	1100	0.5626	0.5449
0.2335	8.05	1200	0.5866	0.5662
0.2031	8.72	1300	0.5574	0.5420
0.1925	9.39	1400	0.5572	0.5297
0.1793	10.07	1500	0.5878	0.5185
0.1652	10.74	1600	0.6173	0.5243
0.1663	11.41	1700	0.5807	0.5133
0.1544	12.08	1800	0.5979	0.5154
0.148	12.75	1900	0.5545	0.4986
0.138	13.42	2000	0.5798	0.4947
0.1353	14.09	2100	0.5670	0.5028
0.1283	14.76	2200	0.5862	0.4957
0.1271	15.43	2300	0.6009	0.4961
0.1108	16.11	2400	0.5873	0.4975
0.1182	16.78	2500	0.6013	0.4893
0.103	17.45	2600	0.6165	0.4898
0.1084	18.12	2700	0.6186	0.4838
0.1014	18.79	2800	0.6122	0.4767
0.1009	19.46	2900	0.5981	0.4793
0.1004	20.13	3000	0.6034	0.4770
0.0922	20.8	3100	0.6127	0.4663
0.09	21.47	3200	0.5967	0.4672
0.0893	22.15	3300	0.6051	0.4611
0.0817	22.82	3400	0.6027	0.4618

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu113
• Datasets 1.18.3
• Tokenizers 0.10.3