ehabnegm/masri-100h-egyptian-tts-enhanced

Masri 100h Enhanced — Egyptian Arabic TTS Corpus

98.2 hours of denoised, loudness-normalized Egyptian Arabic speech with transcripts verified by three independent ASR systems.

This is a production rebuild of two source corpora. Every clip was re-processed from the original audio: noise-suppressed with DeepFilterNet3, normalized to broadcast loudness, and re-transcribed from scratch by two open models plus a commercial arbiter — then kept only where they agree.

---

At a glance

Audio	98.20 hours · 24,470 clips · 327 source videos
Training-ready (tier A+B)	24,445 clips / 98.11 h — 99.9%
Format	WAV PCM 16-bit · 24 kHz mono, embedded in Parquet
Loudness	EBU R128 → −23 LUFS, peak ≤ −1 dBFS
Denoising	DeepFilterNet3 (48 kHz internal, resampled 24k→48k→24k)
Transcripts	3-way consensus, undiacritized, punctuated
Words / vocabulary	897,505 tokens · 102,689 unique surface forms
Clip length	2.51 – 19.00 s (median 17.0 s)
Size	14.96 GB (44 Parquet shards)
License	CC BY-NC 4.0

---

What makes this different

Most machine-transcribed speech corpora ship whatever a single ASR model produced. A single model cannot detect its own hallucinations, so wrong text silently becomes ground truth — and for TTS, text that does not match the audio teaches the model the wrong grapheme→phoneme mapping.

Here every clip was transcribed by two independent architectures, and every disagreement was arbitrated by a third commercial system:

Model	Role	Egyptian WER
oddadmix/cohere-transcribe-arabic-07-2026-dialectal-v2 (2.07B)	primary	0.265
oddadmix/whisper-large-v3-turbo-arabic-dialectal-v2 (809M)	independent check	0.245
Soniox stt-async-v5	arbiter on disagreements only	commercial

Agreement between the two open models: 96.7% (character error rate ≤ 0.15 after Arabic normalization). Median CER between them is 0.0228.

Dialect fidelity

Soniox is the most accurate on content words (proper nouns, numbers, technical terms) but normalizes spelling toward Modern Standard Arabic — هيستمر→سيستمر, كتير→كثير. For an Egyptian TTS corpus that is harmful, so Soniox is used only as a deciding vote; the stored text always keeps the dialect-faithful surface form from the open models whenever they are backed.

---

Quality tiers

Tier	Clips	%	Rule	train_ok
A	23,672	96.7%	cohere == whisper	✅
B	773	3.2%	they differed; Soniox backed one of them	✅
C	25	0.1%	all three differ, or no arbiter	❌

Tier B breakdown: 644 clips where Soniox backed whisper, 129 where it backed cohere.

Filter on train_ok == True for the clean 24,445-clip training set. Tier C is retained for transparency, not deleted — you can inspect exactly what the pipeline was unsure about.

---

Schema

Column	Type	Description
id	string	<video_id>_<index>
audio	Audio	24 kHz mono 16-bit WAV, enhanced + loudness-normalized
text	string	Final transcript — use this
duration	float32	seconds
tier	string	A, B, or C
agreement	string	which systems agreed
train_ok	bool	passes all quality gates
n_words	int32	word count of the final text
cer_cohere_whisper	float32	CER between the two open models
text_cohere	string	raw cohere-v2 output
text_whisper	string	raw whisper-turbo-v2 output
text_soniox	string	Soniox output (disagreements only, else null)
video_id	string	source video
source	string	3shwa_7akawi or noselleel

All three raw transcripts are kept so you can re-derive the consensus with your own thresholds.

---

Quickstart

from datasets import load_dataset

ds = load_dataset("ehabnegm/masri-100h-egyptian-tts-enhanced", split="train")
clean = ds.filter(lambda r: r["train_ok"])       # 24,445 clips / 98.11 h
print(clean[0]["text"], clean[0]["audio"]["sampling_rate"])

Stream it instead of downloading 15 GB:

ds = load_dataset("ehabnegm/masri-100h-egyptian-tts-enhanced", split="train", streaming=True)
for r in ds.take(3):
    print(r["id"], r["tier"], r["text"][:60])

Highest-confidence subset only:

premium = ds.filter(lambda r: r["tier"] == "A" and r["cer_cohere_whisper"] <= 0.05)

---

Processing pipeline

1. Source — 24,470 clips from 100-hour-Egyption-dataset-single-speaker (3shwa + 7akawi, 15,653 clips) and noselleel-egyptian-tts (8,817 clips). Zero ID collisions; audio bytes verified against the Hub before processing.
2. Enhance — DeepFilterNet3 noise suppression. Audio is 24k→48k for the model, then back to 24k.
3. Normalize — EBU R128 integrated loudness to −23 LUFS, then peak-limited to −1 dBFS.
4. Transcribe ×2 — cohere-v2 and whisper-turbo-v2, independently, on the enhanced audio (so text matches exactly the audio that ships), language="ar", punctuation enabled.
5. Compare — Arabic-normalized CER (diacritics stripped; أإآ→ا, ى→ي, ة→ه, ؤ→و, ئ→ي, tatweel removed).
6. Arbitrate — Soniox stt-async-v5 on the 798 disagreements.
7. Tier + gate — duration 1–20 s, ≥2 words, tier A/B → train_ok.

---

Limitations

• Transcripts remain machine-generated. Three-way agreement lowers the error rate; it does not eliminate it. Systematic errors that all models share will survive consensus.
• Not human-verified. No native speaker has reviewed the text.
• Denoising is lossy. DFN3 suppresses noise but can attenuate quiet consonants and breath. The unprocessed originals remain in the two source datasets.
• Punctuation is model-generated — useful as prosody cues, not authoritative.
• Speaker uniformity is not acoustically verified. The 3shwa/7akawi portion is single-narrator by construction; the noselleel portion was segmented in multi-voice mode and may contain guests. Use the source and video_id columns to filter.
• Duration is bunched at the ceiling — the source segmenter capped utterances near 19 s.

---

Ethics & license

Audio derives from publicly available YouTube content. Released CC BY-NC 4.0 for research and non-commercial use. This release grants no rights in the underlying recordings, nor to any speaker's voice or likeness. Obtain permission from the original creators before commercial use or publishing a voice model trained on this data, and disclose synthesized speech as synthetic.

Rights holders: open a discussion here and material will be removed promptly.

---

Citation

@misc{negm_masri100h_enhanced_2026,
  title        = {Masri 100h Enhanced: A Denoised, Consensus-Verified Egyptian Arabic TTS Corpus},
  author       = {Negm, Ehab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ehabnegm/masri-100h-egyptian-tts-enhanced}},
  note         = {98.2 hours, 24,470 clips, 24 kHz, DeepFilterNet3, 3-way ASR consensus}
}

Derived from ehabnegm/100-hour-Egyption-dataset-single-speaker and ehabnegm/noselleel-egyptian-tts.