# Model Card for WhisperLiveSubs
This model is a fine-tuned version of OpenAI's Whisper model on the Common Voice dataset for Urdu speech recognition. It is optimized for transcribing Urdu language audio.

### Model Description
This model is a small variant of the Whisper model fine-tuned on the Common Voice dataset for the Urdu language. It is intended for automatic speech recognition (ASR) tasks and performs well in transcribing Urdu speech.
- **Developed by:** codewithdark
- **Model type:** Whisper-based model for ASR
- **Language(s) (NLP):** Urdu (ur)
- **License:** Apache 2.0
- **Finetuned from model :** openai/whisper-small

## Uses
### Direct Use
This model can be used directly for transcribing Urdu audio into text. It is suitable for applications such as:
- Voice-to-text transcription services
- Captioning Urdu language videos
- Speech analytics in Urdu

### Out-of-Scope Use
The model may not perform well for:
- Non-Urdu languages
- Extremely noisy environments
- Very long audio sequences without segmentation

## How to Get Started with the Model
Use the code below to get started with the model.

```python
from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("codewithdark/WhisperLiveSubs")
model = WhisperForConditionalGeneration.from_pretrained("codewithdark/WhisperLiveSubs")

# Your transcription code here
```

### Training Data
The model was fine-tuned on the Mozilla Common Voice dataset, specifically the Urdu subset. The dataset consists of approximately 141 hr of transcribed Urdu speech.

#### Preprocessing
The audio was resampled to 16kHz, and text was tokenized using the Whisper tokenizer configured for Urdu.

#### Training Hyperparameters
- **Training regime:** Mixed precision (fp16)
- **Batch size:** 8
- **Gradient accumulation steps:** 2
- **Learning rate:** 1e-5
- **Max steps:** 4000

#### Metrics
Word Error Rate (WER) was the primary metric used to evaluate the model's performance.

### Results

- **Training Loss:** 0.2005
- **Validation Loss:** 0.5342
- **WER:** 51.06

*This is my first time fine-tuning this model. Don't worry about the current performance;
improvements can be made to enhance the model's accuracy and reduce the WER.*

- **Hardware Type:** P100 GPU
- **Hours used:** 10 hr
- **Cloud Provider:** Kaggle
- **Compute Region:** PK

### Model Architecture and Objective
The WhisperLiveSubs model is based on the Whisper architecture, designed for automatic speech recognition.

#### Software
- **Framework:** PyTorch
- **Transformers Version:** 

#### Summary
The model demonstrates acceptable performance for Urdu transcription, but there is room for improvement in terms of WER, especially in noisy conditions or with diverse accents.

## Model Card Contact
For inquiries, please contact codewithdark90@gmail.com

@Codewithdark. (2024). WhisperLiveSubs: An Urdu Automatic Speech Recognition Model. Retrieved from https://huggingface.co/codewithdark/WhisperLiveSubs