wav2vec2-urdu / README.md
kingabzpro's picture
Update README.md
f92696a
metadata
language:
  - ur
license: apache-2.0
tags:
  - automatic-speech-recognition
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_7_0
metrics:
  - wer
  - cer
model-index:
  - name: wav2vec2-urdu
    results:
      - task:
          type: automatic-speech-recognition
          name: Speech Recognition
        dataset:
          type: mozilla-foundation/common_voice_7_0
          name: Common Voice ur
          args: ur
        metrics:
          - type: wer
            value: 52.4
            name: Test WER With LM
            args:
              - learning_rate: 0.0003
              - train_batch_size: 64
              - eval_batch_size: 8
              - seed: 42
              - gradient_accumulation_steps: 2
              - total_train_batch_size: 128
              - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
              - lr_scheduler_type: linear
              - lr_scheduler_warmup_steps: 100
              - num_epochs: 100
              - mixed_precision_training: Native AMP
          - type: cer
            value: 26.46
            name: Test CER With LM
            args:
              - learning_rate: 0.0003
              - train_batch_size: 64
              - eval_batch_size: 8
              - seed: 42
              - gradient_accumulation_steps: 2
              - total_train_batch_size: 128
              - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
              - lr_scheduler_type: linear
              - lr_scheduler_warmup_steps: 100
              - num_epochs: 100
              - mixed_precision_training: Native AMP
          - type: wer
            value: 45.63
            name: Test WER LM CV8
          - type: cer
            value: 20.45
            name: Test CER LM CV8

wav2vec2-large-xls-r-300m-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: 0.5747
  • Cer: 0.3268

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: 64
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 128
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 100
  • num_epochs: 100
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
4.3054 16.67 50 9.0055 0.8306 0.4869
2.0629 33.33 100 9.5849 0.6061 0.3414
0.8966 50.0 150 4.8686 0.6052 0.3426
0.4197 66.67 200 12.3261 0.5817 0.3370
0.294 83.33 250 11.9653 0.5712 0.3328
0.2329 100.0 300 7.6846 0.5747 0.3268

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0