MohamedGomaa30/Egyptian-Speech-Clean-MGB3

🏛️ Dataset Card for MGB3-Egyptian-Clean

Dataset Summary

This dataset is a refined and enhanced version of the MGB-3 (Multi-Genre Broadcast) corpus, specifically focused on the Egyptian Arabic dialect. It has been meticulously preprocessed to be "TTS-ready" by combining advanced deep-learning denoising with custom linguistic text normalization.

🛠️ Preprocessing Pipeline

To ensure the highest quality for generative speech tasks (like VITS or MMS finetuning), the following steps were applied to all 6500 samples:

1. Audio Enhancement (DeepFilterNet 3)

We utilized DeepFilterNet 3 to perform high-speed, high-fidelity noise suppression. This stage removes background hiss, hum, and non-stationary noise while preserving the natural characteristics of the Egyptian speaker's voice.

• Target Sampling Rate: 16,000 Hz
• Denoising Engine: Deep signal filtering with ERB (Equivalent Rectangular Bandwidth) masking.

2. Egyptian Text Normalization (Linguistic)

Standard Arabic NLP tools often fail on dialect-specific numbers and expressions. We implemented a custom Egyptian Arabic Normalizer to convert non-verbal tokens into spoken "Masri" words. This is critical for Text-to-Speech (TTS) consistency.

Key Features of the Normalizer:

• Numbers to Words: Converts digits to Egyptian forms (e.g., 3 → "تلاتة" instead of "ثلاثة").
• Currency: Supports EGP, USD, and EUR with Egyptian plural rules (e.g., "$5" → "خمسة دولار").
• Time & Dates: Normalizes "10:30" to "عشرة ونص" and dates to Egyptian ordinal months.
• Mathematical Symbols: Handles percentages ("في المية") and ranges ("من... لحد").

---

📜 Technical Implementation (Code Snippets)

Text Normalization Logic

The normalization uses a rule-based engine to ensure that the "written" text matches exactly how an Egyptian person would "speak" the content.

# Example of the normalization flow used:
text = "السعر 10.50 ج الساعة 10:30"
normalized = normalize_text_for_tts_egyptian(text)
# Output: "السعر عشرة جنيه وخمسين قرش الساعة عشرة ونص"

Audio Cleaning Loop

from df.enhance import enhance, init_df
model, df_state, _ = init_df("DeepFilterNet3")

def process_batch(batch):
    # Audio Denoising
    audio_tensor = torch.from_numpy(batch["audio"]["array"]).unsqueeze(0)
    enhanced = enhance(model, df_state, audio_tensor)
    
    # Text Normalization
    clean_text = normalize_text_for_tts_egyptian(batch["text"])
    return {"text": clean_text, "audio": enhanced.numpy()}

---

📂 Dataset Schema

Each row in the dataset contains:

• text: The normalized Egyptian Arabic transcript (speech-to-word).
• original_audio: The raw MGB-3 audio segment.
• separated_target_audio: The cleaned, denoised 16kHz audio optimized for training.

🏷️ Tags & Metadata

• Language: ar-EG (arz)
• Task: text-to-speech, automatic-speech-recognition
• Dialect: Egyptian
• Processing: Denoised, Normalized

---

📚 Citation & Attribution

1. The Source Repository (Mohamed Rashad)

Since this dataset is a cleaned version of the data hosted by Mohamed Rashad, it is important to credit the initial work of collecting and hosting the Arabic MGB-3 segments:

Mohamed Rashad. (2023). MGB-3 Arabic Dataset for Speech Recognition and Dialect Identification. Available at: https://huggingface.co/datasets/MohamedRashad/MGB-3-Arabic

2. The Original MGB-3 Challenge (Academic Citation)

The MGB-3 dataset was originally released as part of the Multi-Genre Broadcast Challenge. The standard citation for the Egyptian dialect portion is:

@article{DBLP:journals/corr/abs-1709-07276,
  author       = {Ahmed M. Ali and
                  Stephan Vogel and
                  Steve Renals},
  title        = {Speech Recognition Challenge in the Wild: Arabic {MGB-3}},
  journal      = {CoRR},
  volume       = {abs/1709.07276},
  year         = {2017},
  url          = {http://arxiv.org/abs/1709.07276},
  eprinttype    = {arXiv},
  eprint       = {1709.07276},
  timestamp    = {Thu, 13 Jun 2024 17:07:28 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1709-07276.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

3. Software Attribution (DeepFilterNet)

@misc{schröter2022deepfilternet2realtimespeechenhancement,
      title={DeepFilterNet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio}, 
      author={Hendrik Schröter and Alberto N. Escalante-B. and Tobias Rosenkranz and Andreas Maier},
      year={2022},
      eprint={2205.05474},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2205.05474}, 
}