wav2vec2-large-xlsr-tamil-commonvoice

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: 0.6145
• Wer: 0.8512

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: 200
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
12.0478	1.05	100	3.3867	1.0
3.2522	2.11	200	3.2770	1.0
3.1689	3.16	300	3.1135	1.0039
2.9278	4.21	400	2.0485	1.3109
1.3592	5.26	500	0.8044	1.0988
0.7472	6.32	600	0.6571	0.9474
0.5842	7.37	700	0.6079	0.9477
0.4831	8.42	800	0.6083	0.9491
0.4259	9.47	900	0.5916	0.8973
0.3817	10.53	1000	0.6070	0.9147
0.338	11.58	1100	0.5873	0.8617
0.3123	12.63	1200	0.5983	0.8844
0.287	13.68	1300	0.6146	0.8988
0.2706	14.74	1400	0.6068	0.8754
0.2505	15.79	1500	0.5996	0.8638
0.2412	16.84	1600	0.6106	0.8481
0.2176	17.89	1700	0.6152	0.8520
0.2255	18.95	1800	0.6150	0.8540
0.216	20.0	1900	0.6145	0.8512

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu102
• Datasets 1.13.3
• Tokenizers 0.10.3