wav2vec2-Malayalam

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.7479
• Wer: 0.9088

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: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.6036	1.5748	100	6.5081	1.0
3.5056	3.1496	200	3.5634	1.0
3.4952	4.7244	300	3.4927	1.0
3.3772	6.2992	400	3.3696	1.0
3.1849	7.8740	500	3.1735	1.0
1.3056	9.4488	600	1.2938	1.1167
0.8162	11.0236	700	0.8301	1.0190
0.6022	12.5984	800	0.7678	0.9929
0.454	14.1732	900	0.7514	0.9832
0.4104	15.7480	1000	0.7168	0.9452
0.3616	17.3228	1100	0.7297	0.9571
0.2951	18.8976	1200	0.6925	0.9555
0.2667	20.4724	1300	0.7254	0.9400
0.2707	22.0472	1400	0.7498	0.9101
0.2263	23.6220	1500	0.7093	0.9120
0.1933	25.1969	1600	0.7396	0.9091
0.2168	26.7717	1700	0.7417	0.9046
0.2112	28.3465	1800	0.7479	0.9088

Framework versions

• Transformers 4.40.0
• Pytorch 2.2.1+cu121
• Datasets 2.19.1.dev0
• Tokenizers 0.19.1