anuragshas/wav2vec2-xls-r-300m-hi-cv9-with-lm

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - HI dataset. It achieves the following results on the evaluation set:

Loss: 0.5164
Wer: 0.3349
Cer: 0.1082

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: 7.5e-05
train_batch_size: 64
eval_batch_size: 64
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
training_steps: 9815
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.9471	8.16	400	3.7109	1.0	1.0
3.274	16.32	800	3.1582	0.9917	0.9573
1.5889	24.48	1200	0.7763	0.6030	0.1990
1.3647	32.65	1600	0.6051	0.5135	0.1687
1.2532	40.81	2000	0.5423	0.4712	0.1539
1.1905	48.97	2400	0.5180	0.4532	0.1490
1.1193	57.14	2800	0.4906	0.4248	0.1393
1.0584	65.3	3200	0.4854	0.4069	0.1332
1.0095	73.46	3600	0.4780	0.3926	0.1287
0.9759	81.63	4000	0.4666	0.3925	0.1269
0.9593	89.79	4400	0.4808	0.3830	0.1247
0.909	97.95	4800	0.4798	0.3765	0.1212
0.8788	106.12	5200	0.4906	0.3608	0.1162
0.8471	114.28	5600	0.4759	0.3604	0.1166
0.8116	122.44	6000	0.5080	0.3627	0.1176
0.7881	130.61	6400	0.4868	0.3489	0.1135
0.766	138.77	6800	0.4955	0.3492	0.1136
0.7333	146.93	7200	0.5019	0.3461	0.1125
0.709	155.1	7600	0.5084	0.3468	0.1117
0.6911	163.26	8000	0.5144	0.3412	0.1106
0.6683	171.42	8400	0.5219	0.3409	0.1117
0.659	179.59	8800	0.5230	0.3376	0.1096
0.6475	187.75	9200	0.5229	0.3398	0.1097
0.6419	195.91	9600	0.5200	0.3337	0.1084

Framework versions

Transformers 4.19.0.dev0
Pytorch 1.11.0+cu102
Datasets 2.1.1.dev0
Tokenizers 0.12.1