HeshamHaroon/arabic-msa-25k-saudi-male-tashkeel

Arabic MSA 25K — Saudi Male (Tashkeel)

25,000 fully-diacritized Arabic MSA text + audio pairs, rendered with a single Saudi male neural voice at 48 kHz / 16-bit PCM, across 10 thematic categories.

Dataset Summary

arabic-msa-25k-saudi-male-tashkeel is a 25,000-clip Modern Standard Arabic (MSA) speech corpus with matching diacritized text (full tashkeel / ḥarakāt). Every clip is synthesized by the single voice ar-SA-HamedNeural (Azure Neural TTS, Saudi Arabic male) at 48 kHz, 16-bit, mono PCM WAV — ~60.5 hours of audio in total.

The text was generated by GPT-4o-mini under strict rules (MSA only, no dialect, full tashkeel, 13–45 words, no Quranic or poetic content), then synthesized by Azure Speech across two independent regional resources for throughput. A rich per-clip metadata record is provided, including diacritized text, stripped (non-diacritized) text, topic category, word and character counts, tashkeel density, and WAV duration.

Supported Tasks & Use Cases

⚠️ This dataset is not a replacement for human-recorded speech corpora. It is a synthetic dataset produced by a commercial neural TTS system. See Considerations below for how that affects downstream training.

Languages

Modern Standard Arabic (MSA), Saudi accent (spoken via the ar-SA voice). All text is fully diacritized (tashkeel): fatḥa, ḍamma, kasra, sukūn, shadda, tanwīn.

text:          تُعَدُّ الطَاقَةُ الشَّمْسِيَّةُ إِحْدَى أَنْظَفِ مَصَادِرِ الطَاقَةِ الْمُتَجَدِّدَةِ …
text_stripped: تعد الطاقة الشمسية إحدى أنظف مصادر الطاقة المتجددة …

Dataset Structure

Files

├── README.md                                 # this card
├── manifest.json                             # aggregate corpus stats
├── metadata.jsonl                            # 25K metadata rows (provided for manual inspection)
└── data/
    ├── train-00000-of-00005.parquet          # rows 0  –  4,999    (~3.5 GB each)
    ├── train-00001-of-00005.parquet          # rows 5,000 – 9,999
    ├── train-00002-of-00005.parquet          # rows 10,000 – 14,999
    ├── train-00003-of-00005.parquet          # rows 15,000 – 19,999
    └── train-00004-of-00005.parquet          # rows 20,000 – 24,999

Each Parquet shard embeds the 48-kHz WAV bytes inline using HuggingFace's Audio feature — there are no external .wav files to fetch. The datasets library decodes them lazily on access.

Data Instances

A decoded row from the dataset:

{
  "audio": {
      "path": "hamed_000042.wav",
      "array": array([...], dtype=float32),  # shape (num_samples,), 48 kHz
      "sampling_rate": 48000
  },
  "id": "hamed_000042",
  "text": "يُعَدُّ الذَّكَاءُ الاِصْطِنَاعِيُّ فَرْعًا مِنْ فُرُوعِ عِلْمِ الْحَاسُوبِ…",
  "text_stripped": "يعد الذكاء الاصطناعي فرعا من فروع علم الحاسوب…",
  "category": "technology_ai",
  "voice": "ar-SA-HamedNeural",
  "gender": "male",
  "word_count": 28,
  "char_count": 182,
  "tashkeel_density": 0.34,
  "audio_duration_s": 8.12,
}

Data Fields

Categories (10 × 2,500 clips each)

Splits

The dataset ships as a single `train` split of 25,000 rows. Downstream users are free to carve out validation / test splits; we recommend stratifying by category to preserve topic balance.

Loading

from datasets import load_dataset, Audio

ds = load_dataset("HeshamHaroon/arabic-msa-25k-saudi-male-tashkeel", split="train")
# Audio feature auto-decodes on access
ds = ds.cast_column("audio", Audio(sampling_rate=48000))

row = ds[0]
print(row["text"][:80])
print(row["audio"]["array"].shape, row["audio"]["sampling_rate"])
# (418560,) 48000   →   ~8.7 s at 48 kHz

# Stream (no local download of the full 20 GB):
ds_stream = load_dataset("HeshamHaroon/arabic-msa-25k-saudi-male-tashkeel",
                         split="train", streaming=True)
for row in ds_stream.take(3):
    print(row["id"], row["category"], row["text"][:60])

Dataset Creation

Motivation

High-quality Arabic speech data is under-represented relative to English. Even within Arabic, Saudi MSA with full tashkeel and controlled topic diversity is rarely available at this scale. The goal of this release is to provide a clean, legally-unencumbered, synthetic speech corpus that:

1. 1.Uses a single consistent voice (enables speaker-conditional studies),
2. 2.Guarantees full tashkeel coverage on every sample (enables diacritization research and ASR that produces diacritized output),
3. 3.Balances topic categories (no news-only bias), and
4. 4.Is large enough (~60 h) to meaningfully fine-tune small-to-mid ASR / TTS.

Generation pipeline

┌─────────────────────────┐
│ Stage 1 — text gen      │  Azure OpenAI gpt-4o-mini
│ 25,000 unique MSA texts │  temperature 0.6, 40 concurrent
│ 10 categories × 2,500   │  strict validators: MSA-only,
│ Full tashkeel           │  tashkeel density, word count
│                         │  dedup by SHA-256 of text
└─────────────┬───────────┘
              │
              ▼
┌─────────────────────────┐
│ Stage 2 — TTS           │  Azure Speech (ar-SA-HamedNeural)
│ 25,000 × WAV 48 kHz     │  80 concurrent across two independent
│ 16-bit mono PCM         │  regional resources; exponential backoff
│                         │  on HTTP 429
└─────────────┬───────────┘
              │
              ▼
┌─────────────────────────┐
│ Stage 3 — manifest      │  Aggregate metadata into manifest.json
│                         │  (duration, tashkeel density, …)
└─────────────────────────┘

Source Data

Validation & Quality Gates

Every text had to pass all of the following before being synthesized:

• —Tashkeel density (observed mean 0.78).
• —Word count ∈ [13, 45] MSA words.
• —No dialect markers — rejection list includes common Saudi / Egyptian / Levantine / Maghrebi tokens.
• —No Quranic signals — rejection on ﴿﴾ markers or بسم الله الرحمن الرحيم, صلى الله عليه وسلم, etc.
• —Latin / digit caps to keep the text purely Arabic.
• —SHA-256 dedup across categories and generation chunks.

Approximately 43 % of raw model outputs were rejected by these validators, which is why the corpus is synthetic but relatively clean.

Annotations

All annotations (category, word_count, tashkeel_density, audio_duration_s) are programmatic — either direct from the generation pipeline or parsed from WAV headers. No human annotation was performed.

Personal and Sensitive Information

None by construction:

• —Texts were freshly generated by an LLM with no conditioning on personal data.
• —Audio is synthesized by a studio voice and contains no real speaker.
• —No user identifiers, locations, or PII appear in the pipeline.

If any PII or sensitive content slips through the generation filter, please open an issue on the dataset page and we will remove it.

Considerations for Using the Data

Known Limitations

• —Synthetic source — text and audio are both model-generated. Models that train solely on this corpus will inherit any systematic biases of GPT-4o-mini (content bias) and Azure ar-SA-HamedNeural (prosodic / phonetic bias).
• —Average word count is 14.6 — slightly below the 20-word intent of the original spec. Downstream users needing longer utterances should concatenate or augment.
• —Saudi-accented MSA, not pan-Arab-neutral MSA. Phonetic realisations of ق / ض / ج reflect the Saudi voice. For other accents, pair with equivalent generation using ar-EG-*, ar-AE-*, ar-JO-*, etc.
• —Single voice — this corpus cannot be used on its own for multi-speaker TTS / ASR tasks that require speaker diversity.
• —No emotional / expressive variation — the Azure neural voice is used in its default register. All clips are in a neutral tone.
• —Tashkeel from a generative model — while density is high (~0.78), the tashkeel itself has not been individually verified by human linguists. For high-stakes linguistic research (e.g. grammatical case ṣabṭ), cross-check with an MSA tashkeel reference tool.
• —No Whisper round-trip WER is included in this release; Azure OpenAI Whisper rate-limits made a 25K round-trip impractical. A partial WER may be appended in a future revision.

Biases

• —Topic bias — the 10 chosen categories (news, science, health, tech, nature, history, commerce, education, culture, daily life) are themselves a design choice. They omit e.g. sports, politics-heavy commentary, legal language, and dialectal casual speech.
• —Register bias — all clips are formal MSA. Real-world Arabic AI use cases (chat, voice assistants) often need dialectal or colloquial data; this corpus is not that.

Other

The rejection filter explicitly refused Quranic, ḥadīth, and poetic text to avoid releasing TTS renderings of sacred or metrical text without appropriate context. If your application specifically requires religious or poetic audio, use a more suitable dedicated corpus.

Licensing

This dataset is released under the [Creative Commons Attribution 4.0 International license (CC-BY-4.0)](https://creativecommons.org/licenses/by/4.0/).

Users must comply with:

• —Azure OpenAI Code of Conduct,
• —Azure AI terms for synthetic speech output,
• —And attribute this dataset + the generation tooling in downstream derivatives.

Citation

If you use this dataset, please cite it as:

@misc{haroon_arabic_msa_25k_saudi_male_2026,
  title        = {Arabic MSA 25K — Saudi Male (Tashkeel)},
  author       = {Hesham Haroon},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/HeshamHaroon/arabic-msa-25k-saudi-male-tashkeel}},
  note         = {25,000 fully-diacritized Arabic MSA clips synthesized with Azure ar-SA-HamedNeural}
}

Contact & Issues

Found an issue, want a re-run on a different voice, need more clips, or concerned about specific content? Please open an issue on the dataset page or reach out on Hugging Face.