metadata

language:
  - ur
license: apache-2.0
base_model: openai/whisper-tiny
tags:
  - generated_from_trainer
datasets:
  - mozilla-foundation/common_voice_17_0
metrics:
  - wer
model-index:
  - name: Whisper Tiny Urdu
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 17.0
          type: mozilla-foundation/common_voice_17_0
          config: ur
          split: None
          args: 'config: ur, split: test'
        metrics:
          - name: Wer
            type: wer
            value: 16.033947800693557
pipeline_tag: automatic-speech-recognition

Whisper Tiny Urdu

This model is a fine-tuned version of openai/whisper-tiny on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set:

Loss: 0.1247
Wer: 16.0339

Model description

Whisper Tiny Urdu ASR Model This Whisper Tiny model has been fine-tuned on the Common Voice 17 dataset, which includes over 55 hours of Urdu speech data. The model was trained twice with different hyperparameters to optimize its performance:

First Training: The model was trained on the training set and evaluated on the test set for 20 epochs. Second Training: The model was retrained on the combined train and validation sets, with the test set used for validation, also for 20 epochs. Despite being the smallest variant in its family, this model achieves state-of-the-art performance for Urdu ASR tasks. It can be used for deployment on small devices, offering an excellent balance between efficiency and accuracy.

Intended Use:

Intended uses & limitations

This model is particularly suited for applications on edge devices with limited computational resources. Additionally, it can be converted to a FasterWhisper model using the CTranslate2 library, allowing for even faster inference on devices with lower processing power.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 4e-05
train_batch_size: 64
eval_batch_size: 64
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 100
training_steps: 3000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0057	10.1351	1500	0.1443	18.1511
0.0005	20.2703	3000	0.1247	16.0339

Framework versions

Transformers 4.42.3
Pytorch 2.1.2
Datasets 2.20.0
Tokenizers 0.19.1