wav2vec2-base-Tamil-large

This model is a fine-tuned version of facebook/wav2vec2-base-960h on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3015
• Wer: 0.3470
• Cer: 0.0610

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: 6
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 3000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
5.2014	2.2472	300	3.2558	1.0	1.0
1.1765	4.4944	600	0.5245	0.7139	0.1482
0.4765	6.7416	900	0.4191	0.6080	0.1133
0.33	8.9888	1200	0.3494	0.4983	0.0910
0.2488	11.2360	1500	0.3163	0.4470	0.0811
0.1951	13.4831	1800	0.3326	0.4179	0.0748
0.1586	15.7303	2100	0.3135	0.3924	0.0702
0.1308	17.9775	2400	0.3070	0.3798	0.0668
0.1111	20.2247	2700	0.2999	0.3618	0.0635
0.0996	22.4719	3000	0.3015	0.3470	0.0610

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 1.18.3
• Tokenizers 0.19.1