Talha/URDU-ASR

output

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2822
• Wer: 0.2423
• Cer: 0.0842

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

I have used dataset other than mozila common voice, thats why for fair evaluation, i do 80:20 split.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 48
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 192
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 15
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
No log	1.0	174	0.9860	3.1257	1.0
No log	2.0	348	0.9404	2.4914	0.9997
No log	3.0	522	0.1889	0.5970	0.5376
No log	4.0	696	0.1428	0.4462	0.4121
No log	5.0	870	0.1211	0.3775	0.3525
1.7	6.0	1044	0.1113	0.3594	0.3264
1.7	7.0	1218	0.1032	0.3354	0.3013
1.7	8.0	1392	0.1005	0.3171	0.2843
1.7	9.0	1566	0.0953	0.3115	0.2717
1.7	10.0	1740	0.0934	0.3058	0.2671
1.7	11.0	1914	0.0926	0.3060	0.2656
0.3585	12.0	2088	0.0899	0.3070	0.2566
0.3585	13.0	2262	0.0888	0.2979	0.2509
0.3585	14.0	2436	0.0868	0.3005	0.2473
0.3585	15.0	2610	0.2822	0.2423	0.0842

Framework versions

• Transformers 4.21.0
• Pytorch 1.12.0
• Datasets 2.4.0
• Tokenizers 0.12.1