Arabic TTS inspection kit: pronunciation and style test lines

Arabic text reaches a voice without short vowels, so an Arabic text-to-speech system has to decide how each word is said before it can say it. That is where most audible errors come from: درس can be darasa (studied) or darrasa (taught), 3 becomes a masculine or feminine number depending on the noun that follows, and a Najdi listener hears at once when قهوة is said with a q. A general listening test rarely catches these, because the voice sounds fluent while getting the words wrong.

This kit is a short, repeatable test for those errors. Bayanat Labs wrote the lines in Modern Standard Arabic and in Saudi Arabic (Najdi and Hijazi lines and a Saudi code-switching line), with the intended reading and what to listen for. Use it to compare TTS vendors or model versions, to check studio takes before a full recording run, and to agree a record format with your data supplier that keeps voice rights explicit. It contains no audio: it is a script, a rubric and a schema.

What is in the kit

Eighteen lines and what each tests

How to inspect a voice with it

1. Write the reading policy first. Decide MSA or dialect, which Saudi variety, which calendar, how dates, month names and abbreviations are read and how English terms are said. TK11 and TK16 have no right answer until you do.
2. Give the voice the plain text only. Production text has no diacritics, so the test is whether the voice picks the right vowels by itself. Never feed it the reference reading.
3. Keep the first output. Generate once per line and style, or fix a seed, so that two systems or two model versions are compared on equal terms.
4. Listen with the reference reading. A native listener of the target variety writes down what was said, marks each tested word pass or fail with an error code, and scores the style 0 to 2.
5. Report by error type, not one score. A voice that fails every number line needs a text-normalization fix; a Najdi voice that says q in قهوة needs different data or a different speaker. Keep the failing outputs and run the same lines after the fix.

Eighteen lines are a smoke test, not a benchmark. Before launch, extend script.jsonl with your own names, products, prices, dates and code-switched phrases in the same format. Our guide to Arabic TTS datasets explains why diacritics, casting by dialect and recording specifications matter for the data itself.

A worked example: one digit, two readings

TK07 gives the voice اشتريت 3 كتب و3 مجلات (“I bought 3 books and 3 magazines”). The correct reading is ثَلَاثَةَ كُتُبٍ and then ثَلَاثَ مَجَلَّاتْ: kitab is masculine, so its number takes the feminine-looking form with ta, and majalla is feminine, so its number does not. The first illustrative record in manifest_example.jsonl shows a take that reads thalathata (ثلاثة) for both. It is logged as:

{"script_id": "TK07", "style_label": "neutral",
 "transcript_as_spoken": "اِشْتَرَيْتُ ثَلَاثَةَ كُتُبٍ وَثَلَاثَةَ مَجَلَّاتْ.",
 "inspection": {"result": "rerecord",
   "errors": [{"code": "NUM_GENDER", "word": "3", "heard": "ثلاثة مجلات"}],
   "style_match": 2},
 "rights": {"asr_training": "granted", "tts_training": "granted",
   "synthetic_voice_of_speaker": "granted", "redistribution": "not_granted", …},
 "illustrative": true}

For a studio take, a reading slip like this means a retake. For a TTS system, the same error on TK07 or TK08 points to the number-expansion step rather than the voice itself.

Keep each voice right separate

The schema gives every recording a rights object with one field per use: speech-recognition training, TTS training, creating a synthetic voice of the speaker, redistribution and commercial use, plus permitted users, exclusivity, expiry and a pointer to the signed consent. Each field is granted, not_granted or not_stated. The second illustrative record passes inspection but has "synthetic_voice_of_speaker": "not_stated": that recording can be used to train a model only within the rights it states, and it cannot become a reusable voice of that speaker until the permission is in writing. Consent to contribute recordings for speech recognition does not by itself permit a synthetic voice.

What this kit is not

• Not audio. The kit is text only. It contains no recordings and no voice.
• Not a sample of a Bayanat dataset. Bayanat Labs wrote the lines for this kit. The speakers, sessions, inspectors and consent IDs in the example records are fictional.
• Not a benchmark. It has no scores and ranks no system. Eighteen lines show where a voice fails; they do not measure how often.
• Not every variety. v0.1 covers MSA and Saudi Arabic (Najdi and Hijazi). A voice for Egyptian, Levantine or Maghrebi users needs lines written natively in those varieties.

Record and inspect Arabic studio speech with Bayanat Voice Studio

Bayanat Voice Studio supplies Arabic studio recordings you can check with this kit:

• A compact studio set for testing. Studio Prompt Speech has 2,000 Arabic prompts, 40 speakers and 2 hours of audio for controlled speech testing and voice prototyping (Studio Prompt Speech).
• Saudi expressive speech from professional voice actors. Recorded by professional voice actors, with task annotation and commercial AI-use rights; style coverage, collection size and a benchmark report are available on request (Saudi Expressive Speech).
• Custom recordings with pronunciation coverage. Single- or multi-speaker recordings with pronunciation coverage, lexicons, expressive styles and brand-voice briefs, with scripts, style labels and recording conditions reviewed before production (Voice Studio).
• Voice rights in writing. Synthetic voice creation needs explicit terms, and brand-voice agreements define permitted users, exclusivity and reuse (how voice permissions work).
• Native listeners for your variety. A team of 50,000+ expert annotators, including vetted native speakers and linguists across 25+ Arabic varieties, who pass dialect screening before working on client data (Arabic speech data services).
• A pilot first. A fixed-scope pilot with gold-standard QA and a quality report, usually within two weeks, with in-region hosting by default (start a pilot).

Cite or reuse

Free to use, adapt and redistribute under CC BY 4.0 with attribution:

Bayanat Labs (2026). Arabic TTS Inspection Kit v0.1. https://www.bayanatlabs.com/research/arabic-tts-inspection-kit

Found an error, or want test lines for another variety or domain? Write to hello@bayanatlabs.com.