Wav2Vec2_xls_r_openslr_Hi_V2

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

• Loss: 0.3184
• Wer: 0.3104
• Cer: 0.0958

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• num_epochs: 12
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
7.1097	0.48	300	0.9965	3.3989	1.0
3.0235	0.96	600	0.3163	1.3183	0.7977
1.1419	1.44	900	0.1913	0.6416	0.5543
0.8242	1.92	1200	0.1608	0.5063	0.4804
0.6876	2.56	1600	0.1387	0.4401	0.4280
0.5868	3.21	2000	0.1249	0.3940	0.3907
0.5285	3.85	2400	0.1200	0.3661	0.3763
0.5	4.49	2800	0.3528	0.3610	0.1136
0.4538	5.13	3200	0.3403	0.3485	0.1086
0.4165	5.77	3600	0.3335	0.3439	0.1062
0.3989	6.41	4000	0.3264	0.3340	0.1036
0.3679	7.05	4400	0.3256	0.3287	0.1013
0.3517	7.69	4800	0.3212	0.3223	0.1002
0.3357	8.33	5200	0.3173	0.3196	0.0986
0.3225	8.97	5600	0.3142	0.3177	0.0985
0.3057	9.62	6000	0.3199	0.3156	0.0975
0.2972	10.26	6400	0.3139	0.3128	0.0967
0.2881	10.9	6800	0.3184	0.3107	0.0957
0.2791	11.54	7200	0.3184	0.3104	0.0958

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0