Languages & dialects
Nabra listens for the language and the dialect itself. You send a schema and the input — no locale, no “Saudi mode”, no “Jordanian mode”. Dialect knowledge only sharpens what the engine already does: numbers, dates, phone formats and currency.
Auto by default
The same request works for a caller in Riyadh, Beirut or London. The response tells you what was heard — and a developer can still pin a language when they know it.
From the script, the speech recogniser, and the words — including mixed Arabic–English sentences.
Levantine or Gulf first; a specific dialect only when the evidence is strong. Otherwise it stays “auto” — never a guess.
“بدي أحجز appointment tomorrow at five” is one sentence, not an error. Values stay in the language they were said in.
Whatever dialect the person speaks, follow-up questions come back in English or in clear, formal Arabic that everyone understands.
curl -X POST https://api.example.com/api/v1/voice/parse \
-H "Authorization: Bearer $API_KEY" \
-F "audio=@voice.webm" \
-F "schema=appointment_booking""language": {
"requested": "auto",
"detected": "ar",
"locale": "ar",
"dialectGroup": "gulf",
"dialect": "auto",
"direction": "rtl",
"mixedLanguage": false,
"source": "detected",
"evidence": ["0551234567", "ابغي", "بكره"]
}Dialect families
Jordan, Palestine, Lebanon and Syria — “بدي”, “بكرا”, “هلّق”, “شو”.
العربية (الأردن)
ar-JOSupportedSomeone says
«بدي أبعث طلب من الجبيهة للشميساني، المستلم أحمد ورقمه 0791234567، والطرد عبارة عن كيس أكل.»
What the engine does with it
Knows places like خلدا · الجاردنز · عبدون · الصويفية
Detection, measured · 20
right 100% of the times it commits
العربية (فلسطين)
ar-PSSupportedSomeone says
«بدي أحجز موعد بكرا الساعة خمسة، اسمي سامر ورقمي 0591234567.»
What the engine does with it
Knows places like رام الله · البيرة · نابلس · الخليل
Detection, measured · 20
right 100% of the times it commits
العربية (لبنان)
ar-LBSupportedSomeone says
«بدي إحجز موعد بكرا الساعة خمسة، اسمي رامي ورقمي 03123456.»
What the engine does with it
Knows places like بيروت · الأشرفية · الحمرا · جونية
Detection, measured · 20
right 100% of the times it commits
العربية (سوريا)
ar-SYSupportedSomeone says
«بدي احجز موعد بكرا الساعة خمسة، اسمي باسل ورقمي 0944123456.»
What the engine does with it
Knows places like الشام · دمشق · المزة · باب توما
Detection, measured · 20
right 100% of the times it commits
Saudi Arabia, the UAE, Kuwait, Qatar, Bahrain and Oman — “أبي/أبغى”, “باجر”, “الحين”, “وش/شنو”.
العربية (السعودية)
ar-SASupportedSomeone says
«أبغى أحجز موعد بكرة الساعة خمسة، اسمي فهد ورقمي 0551234567.»
What the engine does with it
Knows places like الرياض · جدة · الدمام · الخبر
Detection, measured · 20
right 100% of the times it commits
العربية (الإمارات)
ar-AESupportedSomeone says
«أبا أحجز موعد باجر الساعة خمس، اسمي مريم ورقمي 0501234567.»
What the engine does with it
Knows places like دبي · أبوظبي · الشارقة · العين
Detection, measured · 20
right 100% of the times it commits
العربية (الكويت)
ar-KWSupportedSomeone says
«أبي أحجز موعد باچر الساعة خمس، اسمي عبدالله ورقمي 55123456.»
What the engine does with it
Knows places like الكويت · حولي · السالمية · الجابرية
Detection, measured · 20
right 100% of the times it commits
العربية (قطر)
ar-QASupportedSomeone says
«أبي أحجز موعد باچر الساعة خمس، اسمي حمد ورقمي 55123456.»
What the engine does with it
Knows places like الدوحة · الوكرة · الريان · لوسيل
Detection, measured · 20
right 100% of the times it commits
العربية (البحرين)
ar-BHSupportedSomeone says
«أبي أحجز موعد باچر الساعة خمس، اسمي علي ورقمي 36123456.»
What the engine does with it
Knows places like المنامة · المحرق · الرفاع · مدينة عيسى
Detection, measured · 20
right 100% of the times it commits
العربية (عُمان)
ar-OMSupportedSomeone says
«بغيت أحجز موعد باكر الساعة خمس، اسمي سالم ورقمي 92123456.»
What the engine does with it
Knows places like مسقط · مطرح · السيب · الخوض
Detection, measured · 20
right 100% of the times it commits
First-class, not an afterthought: spoken numbers, “half past five”, “the day after tomorrow”, US and UK phone formats.
English (US)
en-USSupportedSomeone says
«I'd like to book an appointment tomorrow at five pm. My name is Sara, my number is 202 555 0143.»
What the engine does with it
Language detected in 100% of 30 English cases
English (UK)
en-GBSupportedSomeone says
«I'd like to book an appointment tomorrow at five pm. My name is Sara, my number is 07700 900123.»
What the engine does with it
Language detected in 100% of 30 English cases
Code-switching mid-sentence, Arabic written in Latin letters around English words, English numbers inside Arabic sentences. Detected as mixed, extracted as one request.
«بدي أحجز appointment بكرا at 5، اسمي Sara ورقمي 0791234567»
Language detected in 100% of 20 mixed cases
Registered with their words and phone formats, but not yet benchmarked. Arabic still works for them through the shared pipeline.
العربية (العراق)
ar-IQExperimentalSomeone says
«أريد أحجز موعد باچر الساعة خمسة، اسمي علي ورقمي 07701234567.»
What the engine does with it
Not benchmarked yet
العربية (مصر)
ar-EGExperimentalSomeone says
«عايز أحجز ميعاد بكره الساعة خمسة، اسمي محمد ورقمي 01012345678.»
What the engine does with it
Not benchmarked yet
العربية (اليمن)
ar-YEExperimentalSomeone says
«اشتي أحجز موعد بكره الساعة خمس، اسمي عبدالله ورقمي 771234567.»
What the engine does with it
Not benchmarked yet
Same meaning, many words
A few of the words the lexicon knows for each dialect. “باكر” means tomorrow in Oman but “early” elsewhere — so it only counts as a date when Gulf speech is detected.
| Dialect | I want | tomorrow | now | what |
|---|---|---|---|---|
| Jordan | بدي | بكرا | هسّا / هلأ | شو |
| Palestine | بدي | بكرا | هلقيت | شو |
| Lebanon | بدي | بكرا | هلّق | شو |
| Syria | بدي | بكرا | هلق | شو |
| Saudi Arabia | أبغى | بكرة | دحين / الحين | وش / إيش |
| United Arab Emirates | أبا / أبي | باجر | الحين | شو |
| Kuwait | أبي | باچر | الحين | شنو |
| Qatar | أبي | باچر | الحين | شنو / وش |
| Bahrain | أبي | باچر | الحين | شنو |
| Oman | بغيت | باكر | الحين | إيش |
| Iraq | أريد | باچر | هسّه | شنو |
| Egypt | عايز | بكره | دلوقتي | إيه |
| Yemen | اشتي | بكره | ذلحين | إيش |
What’s measured — and what isn’t
Deterministic, no model: language detected correctly in all 250 synthetic text cases; the dialect family in 97%. The exact dialect is named in about 40% of short requests — and when it is named, it has been right every time. When unsure, it says “auto”.
Qwen 2.5 7B and Aya Expanse 8B are being benchmarked on the same 250 cases. Results are published in the docs when the runs complete — no number is shown here before it’s measured.
Not measured. The dataset is text; per-dialect speech accuracy needs consented recordings in each dialect.
Open the live demo and speak the way you normally do — English, Arabic, or both.