🏛️ Dataset Card for Egy-Speech-DeepClean_v0

Dataset Summary

Egy-Speech-DeepClean_v0 is a high-fidelity, refined version of Egyptian Arabic speech data, specifically engineered for generative speech tasks like Text-to-Speech (TTS) and Automatic Speech Recognition (ASR). This dataset addresses the common "noise" and "non-verbal" issues found in raw broadcast data by applying state-of-the-art deep-learning denoising and dialect-specific linguistic normalization.

---

🛠️ Preprocessing Pipeline

To transform raw recordings into training-ready assets, the dataset underwent a two-stage cleaning pipeline:

1. Audio Enhancement (DeepFilterNet 3)

We utilized DeepFilterNet 3 to extract the speaker's voice from complex backgrounds. This stage is crucial for removing studio hiss, background music, and street noise without introducing the metallic artifacts common in traditional spectral subtraction.

• Target Sampling Rate: 16,000 Hz (Mono)
• Technique: Perceptual loss-based filtering with ERB masking.

2. Egyptian Text Normalization (Linguistic)

Egyptian Arabic is often written in Modern Standard Arabic (MSA) style but spoken differently. This dataset features a "Spoken-Form" normalization to ensure the text matches the audio exactly.

Feature	Input (Written)	Output (Spoken/Normalized)
Numbers	5 أشخاص	خمسة أشخاص
Currency	100 ج.م	مية جنيه
Time	3:30	تلاتة ونص
Symbols	50%	خمسين في المية

---

📜 Technical Implementation

Text Normalization Logic

The normalization uses a custom rule-based engine to handle the nuances of the "Masri" dialect, ensuring that digits and abbreviations are expanded into their colloquial phonetic equivalents.

# Example of the normalization flow:
raw_text = "السعر 10.50 ج.م الساعة 12:00"
clean_text = normalize_egyptian_arabic(raw_text)
# Output: "السعر عشرة جنيه وخمسين قرش الساعة اتناشر"

Audio Cleaning Loop

The following logic was applied to ensure every sample meets the "DeepClean" standard:

from df.enhance import enhance, init_df
import torch

# Initialize DeepFilterNet 3
model, df_state, _ = init_df("DeepFilterNet3")

def deep_clean_sample(batch):
    # 1. Denoise Audio
    audio_in = torch.from_numpy(batch["audio"]["array"]).unsqueeze(0)
    enhanced_audio = enhance(model, df_state, audio_in)
    
    # 2. Normalize Text to 'Masri' Spoken Form
    normalized_text = normalize_egyptian_arabic(batch["text"])
    
    return {"text": normalized_text, "audio": enhanced_audio.numpy()}

---

📂 Dataset Schema

Each entry in the dataset consists of the following fields:

• text: The cleaned, normalized Egyptian transcript (ready for TTS input).
• audio: The 16kHz denoised audio segment (optimized for training).
• original_text: The raw, unnormalized transcript for reference.
• duration: The length of the audio segment in seconds.

---

🏷️ Tags & Metadata

• Language: ar-EG (Egyptian Arabic)
• Task: text-to-speech, speech-recognition
• Processing: DeepFilterNet3, Text-Normalized
• License: (Specify your license, e.g., MIT or CC-BY-4.0)

---

📚 Citation & Attribution

1. The Source

This dataset is a refined version of the Egyptian cleaned dataset curated by MAdel121. We acknowledge his significant contribution to the Egyptian AI community.

• Original Data: MAdel121/arabic-egy-cleaned

2. DeepFilterNet Citation

@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}, 
}