README.md

metadata

language:
  - ur
license: apache-2.0
tags:
  - automatic-speech-recognition
  - robust-speech-event
datasets:
  - common_voice
metrics:
  - wer
  - cer
model-index:
  - name: wav2vec2-large-xlsr-53-urdu
    results:
      - task:
          type: automatic-speech-recognition
          name: Urdu Speech Recognition
        dataset:
          type: common_voice
          name: Urdu
          args: ur
        metrics:
          - type: wer
            value: 59.2
            name: Test WER
            args:
              - 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
          - type: cer
            value: 32.9
            name: Test CER
            args:
              - 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

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: 0.5921
• Cer: 0.3288

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 Urdu-60 checkpoint and finetune the wav2vwc2 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	Wer	Cer
13.83	8.33	100	0.6611	0.3639
1.0144	16.67	200	0.6498	0.3731
0.5801	25.0	300	0.6454	0.3767
0.3344	33.33	400	0.6349	0.3548
0.1606	41.67	500	0.6105	0.3348
0.0974	50.0	600	0.5921	0.3288

Framework versions

• Transformers 4.15.0
• Pytorch 1.10.0+cu111
• Datasets 1.17.0
• Tokenizers 0.10.3