kingabzpro's picture
eval
0f8f63e
metadata
language:
  - ur
license: apache-2.0
tags:
  - automatic-speech-recognition
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
metrics:
  - wer
  - cer
model-index:
  - name: wav2vec2-urdu-V8-Abid
    results:
      - task:
          type: automatic-speech-recognition
          name: Speech Recognition
        dataset:
          type: mozilla-foundation/common_voice_8_0
          name: Common Voice ur
          args: ur
        metrics:
          - type: wer
            value: 44.63
            name: Test WER
            args:
              - learning_rate: 0.000075
              - 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 AMPP
          - type: cer
            value: 18.82
            name: Test CER
            args:
              - learning_rate: 0.000075
              - 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 AMPP

wav2vec2-60-Urdu-V8

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:

  • Loss: 11.4832
  • Wer: 0.5729
  • Cer: 0.3170

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 7.5e-05
  • 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
19.671 8.33 100 7.7671 0.8795 0.4492
2.085 16.67 200 9.2759 0.6201 0.3320
0.6633 25.0 300 8.7025 0.5738 0.3104
0.388 33.33 400 10.2286 0.5852 0.3128
0.2822 41.67 500 11.1953 0.5738 0.3174
0.2293 50.0 600 11.4832 0.5729 0.3170

Framework versions

  • Transformers 4.16.2
  • Pytorch 1.10.0+cu111
  • Datasets 1.18.3
  • Tokenizers 0.11.0