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: 57.7
            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: 33.8
            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:

• Loss: 11.4593
• Wer: 0.5772
• Cer: 0.3384

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 checkpoint and finetune the XLSR model.

Training and evaluation data

Trained on Harveenchadha/vakyansh-wav2vec2-urdu-urm-60 due to lesser number of samples. Persian and Urdu are quite similar.

Training procedure

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
13.2136	8.33	100	9.5424	0.7672	0.4381
2.6996	16.67	200	8.4317	0.6661	0.3620
1.371	25.0	300	9.5518	0.6443	0.3701
0.639	33.33	400	9.4132	0.6129	0.3609
0.4452	41.67	500	10.8330	0.5920	0.3473
0.3233	50.0	600	11.4593	0.5772	0.3384

Framework versions

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