wav2vec2-large-xlsr-53-urdu
This model is a fine-tuned version of Harveenchadha/vakyansh-wav2vec2-urdu-urm-60 on the common_voice dataset. It achieves the following results on the evaluation set:
- Wer: 0.5913
- Cer: 0.3310
Model description
The training and valid dataset is 0.58 hours. It was hard to train any model on lower number of so I decided to take vakyansh-wav2vec2-urdu-urm-60 checkpoint and finetune the wav2vec2 model.
Training procedure
Trained on Harveenchadha/vakyansh-wav2vec2-urdu-urm-60 due to lesser number of samples.
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: 50
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
---|---|---|---|---|---|
12.6045 | 8.33 | 100 | 8.4997 | 0.6978 | 0.3923 |
1.3367 | 16.67 | 200 | 5.0015 | 0.6515 | 0.3556 |
0.5344 | 25.0 | 300 | 9.3687 | 0.6393 | 0.3625 |
0.2922 | 33.33 | 400 | 9.2381 | 0.6236 | 0.3432 |
0.1867 | 41.67 | 500 | 6.2150 | 0.6035 | 0.3448 |
0.1166 | 50.0 | 600 | 6.4496 | 0.5913 | 0.3310 |
Framework versions
- Transformers 4.15.0
- Pytorch 1.10.0+cu111
- Datasets 1.17.0
- Tokenizers 0.10.3
- Downloads last month
- 28
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Dataset used to train kingabzpro/wav2vec2-60-urdu
Evaluation results
- Test WER on Common Voice urself-reported59.100
- Test CER on Common Voice urself-reported33.100