---
language: ar
license: apache-2.0
tags:
- whisper
- automatic-speech-recognition
- asr
- audio
- arabic
- egyptian-arabic
datasets:
- MAdel121/arabic-egy-cleaned
metrics:
- wer
- cer
base_model: openai/whisper-medium
pipeline_tag: automatic-speech-recognition
library_name: transformers
model-index:
- name: whisper-medium-egy
  results:
  - task:
      type: automatic-speech-recognition
      name: Speech Recognition
    dataset:
      name: MAdel121/arabic-egy-cleaned (validation split)
      type: MAdel121/arabic-egy-cleaned
      config: ar
      split: validation
    metrics:
    - name: WER
      type: wer
      value: 18.029990439289488
    - name: CER
      type: cer
      value: 13.375029793807732
---

# Whisper Medium Egyptian Arabic (whisper-medium-egy)

This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on a custom dataset of 72 hours of Egyptian Arabic speech. It's designed for Automatic Speech Recognition (ASR) for the Egyptian Arabic dialect.

## Model Description

*   **Base Model:** `openai/whisper-medium`
*   **Language:** Arabic (ar), specifically focused on Egyptian dialect (arz)
*   **Fine-tuning Dataset:** `MAdel121/arabic-egy-cleaned` (approx. 72 hours)
*   **Total Training Steps:** 7299
*   **Epochs:** 10

## Intended Uses & Limitations

This model is intended for transcribing speech in Egyptian Arabic.

**Intended Use:**
*   Automatic transcription of audio recordings and live speech in Egyptian Arabic.
*   Assisting with content creation, subtitling, and voice-controlled applications for Egyptian Arabic speakers.

**Limitations:**
*   Performance may degrade in highly noisy environments or with very strong, non-Egyptian accents.
*   The model was fine-tuned on a specific dataset; its performance on significantly different domains or audio characteristics might vary.
*   The training data primarily consists of [describe your dataset sources/domains if possible, e.g., "YouTube videos", "audiobooks", "scripted conversations"]. Performance might be better on similar types of audio.

## How to Use

You can use this model with the `transformers` library and the `pipeline` interface for ease of use.

```python
from transformers import pipeline
import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"

pipe = pipeline(
  "automatic-speech-recognition",
  model="YOUR_HF_USERNAME/whisper-medium-egy", # Replace YOUR_HF_USERNAME with your Hugging Face username
  device=device
)

# Example with a local audio file
# audio_file = "path/to/your/egyptian_arabic_audio.wav"
# transcription = pipe(audio_file, generate_kwargs={"language": "arabic"})["text"]
# print(transcription)

# Example with a Hugging Face dataset audio sample
# from datasets import load_dataset
# ds = load_dataset("MAdel121/arabic-egy-cleaned", "ar", split="validation") # Or your test split
# sample = ds[0]["audio"] # Make sure your dataset has an "audio" column
# result = pipe(sample.copy(), generate_kwargs={"language": "arabic"})
# print(result["text"])
```
Make sure to replace `"YOUR_HF_USERNAME/whisper-medium-egy"` with the actual model ID after uploading. The `generate_kwargs={"language": "arabic"}` is important for Whisper models to ensure correct tokenization and transcription for the target language.

## Training Data

The model was fine-tuned on the `MAdel121/arabic-egy-cleaned` dataset available on the Hugging Face Hub. This dataset contains approximately 72 hours of Egyptian Arabic audio paired with transcripts.

## Training Procedure

The model was trained using the `transformers` library. The fine-tuning process involved the following key hyperparameters:

*   **Base Model:** `openai/whisper-medium`
*   **Optimizer:** AdamW
*   **Learning Rate:** 1e-5 (0.00001)
*   **Warmup Steps:** 1000
*   **Weight Decay:** 0.05
*   **Gradient Accumulation Factor:** 2
*   **Batch Size (loader_batch_size):** 8 (effective batch size would be 8 * 2 = 16)
*   **Number of Epochs:** 10
*   **Max Grad Norm:** 5
*   **Augmentations Used:**
    *   `use_drop_freq`: true
    *   `use_drop_chunk`: true
    *   `use_drop_bit_resolution`: true
    *   Other augmentations like `use_add_noise`, `use_speed_perturb`, `use_pitch_shift`, `use_add_reverb`, `use_codec_augment`, `use_gain` were set to `false` 
*   **Task:** transcribe
*   **Language:** ar
*   **Seed:** 1986

Training was done on 1x A100 (80GB) on Modal Labs

The training was managed and tracked using Weights & Biases under the project `whisper-medium-egyptian-arabic` with resume ID `r3sz4v27`.

## Training Code

Can be found on [Github here](https://github.com/moadel321/Fine-tuning-whisper-on-Modal-Labs-with-speech-brain-augmentations-/blob/c85312785faa2b927cbc217fe43acb8ed660d2ee/train_whisper_modal.py)

## Weights & Biases 

Run can be found here : https://wandb.ai/m-adelomar1/whisper-medium-egyptian-arabic/

## Evaluation Results

The model was evaluated on the `validation` split of the `MAdel121/arabic-egy-cleaned` dataset.

*   **Word Error Rate (WER):** 18.03%
*   **Character Error Rate (CER):** 13.38%

These metrics indicate the performance of the model on the validation set. Lower values are better.

### BibTeX Citation

```bibtex
@misc{madel_2025_whisper_medium_egy,
  author    = Madel
  title     = {Whisper Medium Fine-tuned for Egyptian Arabic},
  year      = {2025},
  publisher = {Hugging Face},
  journal   = {Hugging Face Hub},
  howpublished = {\\url{https://huggingface.co/MAdel121/whisper-medium-egy}} // Replace with actual URL
}
```