MohamedGomaa30/Synthetic-Egy-Speech-Dataset

Synthetic Egyptian Speech Dataset

A curated dataset of 1000 Egyptian Arabic speech samples — the best audio selected across 4 TTS models for each prompt, with transcription text and quality metadata.

Dataset Description

Each entry contains:

• id: Unique prompt identifier (e.g., egy_0001)
• text: Egyptian Arabic transcription text
• audio_path: Path to the best-selected .wav audio file
• model: TTS model that produced the best audio (lahgtna, chatterbox_egyptian, egtts_v01, or habibi)
• mcd: Mel Cepstral Distortion score (spectral similarity to original)

Selection Criteria

For each prompt, the best model was selected by:

1. Lowest WER degradation (among accepted samples)
2. Tiebreak: highest MOS → lowest CER → alphabetical
3. Fallback: lowest synth WER (if no accepted entries)

Model Distribution

Model	Selected	Accept Rate
lahgtna	~600	40.0%
chatterbox_egyptian	~300	30.1%
egtts_v01	~90	18.3%
habibi	~10	1.0%

Source

Generated by the Egyptian Arabic Synthetic Speech Data Pipeline from MohamedGomaa30/Egyptian-Speech-Audio-Text-1k.

Citation

@software{egyptian_tts_pipeline,
  title={Egyptian Arabic Synthetic Speech Data Pipeline},
  author={Mohamed Gomaa Elsafy},
  year={2026},
  url={https://github.com/Mohamed-Gomaa30/Synthetic-Speech-Data-Pipeline}
}