urdumodel

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: 0.4939
Wer: 0.3698
Cer: 0.1465

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

For training 95 hours of audio data is used. For evaluation test set of common voice 10.0 is used.

Training procedure

All the code is available here https://github.com/talhaanwarch/Urdu-ASR

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 24
eval_batch_size: 24
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 96
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 20
mixed_precision_training: Native AMP

Model score on test

When I train I got different WER and CER score on test set, but when I tested locally I got WER of 0.27 without language model and 0.22 with language model.

Framework versions

Transformers 4.21.1
Pytorch 1.12.0
Datasets 2.4.0
Tokenizers 0.12.1

Talha
/

urdumodel