mohammedaly22/lahgtna-levantine-tts

🎙️ Lahgtna Levantine TTS

Synthetic Levantine Arabic + English Code-Switching speech dataset. Generated using Lahgtna-OmniVoice, a fine-tuned zero-shot TTS model for Levantine Arabic dialect.

---

📊 Dataset Statistics

Metric	Value
Total utterances	50,000
Total speakers	10 (5 male, 5 female)
Pure Levantine Arabic	44,154 utterances
Code-switching (AR+EN)	5,846 utterances
Sampling rate	24,000 Hz
Estimated total duration	~66.8 hours

Per-Speaker Breakdown

Speaker ID	Name	Gender	Utterances	Pure AR	Code-Switch	Est. Hours
spk_01_male	Badr	male	5,000	4,429	571	6.86h
spk_02_male	Mohamed	male	5,000	4,433	567	6.71h
spk_03_male	Saad	male	5,000	4,419	581	6.28h
spk_04_male	Rami	male	5,000	4,393	607	7.05h
spk_05_male	Fadi	male	5,000	4,428	572	6.42h
spk_06_female	Amina	female	5,000	4,379	621	5.97h
spk_07_female	Fatma	female	5,000	4,420	580	5.88h
spk_08_female	Lamyaa	female	5,000	4,426	574	7.36h
spk_09_female	Mona	female	5,000	4,428	572	7.53h
spk_10_female	Haneen	female	5,000	4,399	601	6.71h

---

📝 Data Collection & Processing

1. Text Data Sources

The 50,000 sentences were collected from:

Source	Type	Count
GU-CLASP Shami Corpus	Real Levantine Arabic (Syrian, Lebanese, Palestinian, Jordanian)	~44,000
Synthetic code-switching templates	Levantine Arabic + English (tech/daily life)	~6,000

The Shami corpus provides authentic dialectal text from four Levantine sub-dialects:

• Syrian (syrian.txt) — 34,491 sentences
• Lebanese (Lebenees.txt) — 9,905 sentences
• Palestinian (Palestinian.txt) — 9,545 sentences
• Jordanian (jordinian.txt) — 6,007 sentences

Code-switching sentences follow natural Levantine-English mixing patterns:

هَلَّق عم أشتغل على the project اللي حكيتلك عنه
والله the meeting كتير important، لازم نحضّر مِنِيح

2. Text Normalization & Partial Diacritization

Before synthesis, each sentence was processed through:

Step 1 — Unicode cleanup: NFC normalization, tatweel removal, alef unification

Step 2 — Number verbalization: Levantine Arabic number words

• 3 كتب → تلاتة كتب
• $50 → خمسين دولار

Step 3 — Partial diacritization on homographs only: The key design decision: instead of full diacritization, we apply diacritics only to ambiguous homographs that could be mispronounced. This makes the model robust to both diacritized and undiacritized input at inference time.

Diacritized homograph examples:

هَلَّق  (now — vs هَلَقَ = he shaved, MSA)
ضَلّ   (remained, Levantine — vs ضَلَّ = went astray, MSA)
مِشْ   (not, Levantine negation)
بِدِّي (I want, Levantine bi-imperfect)

Step 4 — ه → ة correction: Levantine Arabic informal writing uses ه where standard orthography uses ة (ta marbuta). A comprehensive rule-based corrector fixes feminine nouns, adjectives, and proper names while preserving genuine ه in verb+pronoun forms and الله compounds:

• هالضحكه الحلوه → هالضحكة الحلوة ✅
• والله → والله (preserved — contains الله) ✅
• فيه، عليه، معه → preserved (pronoun suffixes) ✅

Step 5 — Levantine lexicon overrides (148 entries in CSV): Common Levantine dialect words get dialect-correct diacritization via an editable CSV file (data/levantine_lexicon.csv) — no code changes needed to add new words.

3. TTS Synthesis — Lahgtna-OmniVoice

Property	Value
Model	oddadmix/lahgtna-omnivoice-v2
Base architecture	OmniVoice (k2-fsa/OmniVoice fine-tune)
Fine-tuning	Levantine Arabic dialect (apc — ISO 639-3)
Generation mode	Zero-shot voice cloning from reference audio
Language code	apc (North Levantine Arabic)
Output sample rate	24,000 Hz
Generation parameters	temperature=0.7, top_p=0.7, repetition_penalty=1.2

Each speaker was cloned from a 5–15 s reference recording of a real Levantine speaker. The 10 speakers were generated in parallel across 4× NVIDIA H100 GPUs using Python multiprocessing, with each GPU handling 2–3 speakers simultaneously.

---

📁 Dataset Structure

train/
  audio       — Audio feature at 24 kHz
  text        — Levantine Arabic transcript (partial diacritics on homographs)
  speaker_id  — e.g. "spk_01_male"
  speaker_name— e.g. "Badr"
  gender      — "male" | "female"
  sentence_type — "pure_levantine" | "code_switching"

---

🔧 Usage

from datasets import load_dataset

ds = load_dataset("mohammedaly22/lahgtna-levantine-tts", split="train")

# Play sample
sample = ds[0]
print(sample["text"])          # transcript
print(sample["speaker_name"]) # e.g. "Badr"
print(sample["sentence_type"]) # "pure_levantine" or "code_switching"

# Audio: sample["audio"]["array"] at 24000 Hz

---

📜 License

CC BY 4.0 — Free to use with attribution.

👤 Author

🔗 Related

• Lahgtna-OmniVoice model
• oddadmix/lahgtna-omnivoice-v2
• GU-CLASP/shami-corpus
• leva-tts training repo