لهجات أصيلة
/ʕ/Dialect-native voices
Multi-dialect speech today, each with native rhythm and vowel color, not a re-accented standard.
Dialect-native Arabic voice - synthesis & recognition
Natiq — an Arabic voice model built around the language as it is spoken.
Mithal builds Arabic speech models for natural text-to-speech and transcription across dialects.
text-to-speechtranscription (ASR)voice cloning from secondsmulti-dialectArabic-first modeling
§ 01 - Listen
Same meaning. Four mother tongues.
Pre-rendered samples — nothing is generated live here.
text to speech
Modern Standard
كان العمل اليوم مزدحمًا جدًا، كان لديّ اجتماع تلو الآخر، ولم أتمكن حتى من أخذ استراحة.
settings
dialect
voice
speaker encoder
cfg scale1.40
dry multiplier0.80
الفصحى
current target
§ 02 - The model
Not a multilingual model with Arabic bolted on. A speech model built around the language's structure.
استنساخ الصوت
/q/Voice cloning
Capture a speaker from a few seconds of reference audio with the speaker encoder.
~6s reference -> matched timbre
التشكيل التلقائي
/h/Diacritization-aware
Reads bare, undiacritized text and infers the vowels the way a native speaker would.
علم -> عِلْم / عَلَم
تبديل لغوي
/sh/Code-switching
Moves between Arabic and English mid-sentence without breaking prosody or pace.
AR + EN, one breath
تحكّم بحثي
/s/Research-grade control
CFG scale, candidates, and sampling controls. Real parameters, not one quality slider.
cfg · candidates · min-p · dry
بنية عربية أولاً
/dh/Arabic-first design
Tokenization, conditioning, and evaluation choices shaped around Arabic rather than inherited from a general multilingual system.
language-specific by design
§ 03 - Platform
A first look at the workspace behind the voice.
Write in Arabic, choose a dialect and voice, then shape the result from one focused interface.
NATIQ V1
أهلًا بيكم، دي أول مرة تسمعوا فيها صوت ناطق باللهجة المصرية.
Try an example
MSA
Egyptian
Emirati
Saudi
Generate
§ 04 - Reading the unwritten
Arabic leaves the vowels out. The same three letters, three different words.
Natiq infers the missing vowels from context, then says the sentence the way a native speaker would.
bare text
علم
عِلْم
ilmknowledge
العِلْمُ نورٌ يهدي صاحبَه إلى الحقيقة.
عَلَم
alama flag
رُفِعَ العَلَمُ عاليًا فوق السارية.
عَلَّمَ
allamahe taught
عَلَّمَ المعلمُ تلاميذَه بصبرٍ وحبّ.
§ 05 - Codeswitching
Arabic and English can share the same utterance without breaking the rhythm of the sentence.
ممكن تعملليforwardللـemail؟
§ 06 - And it listens
The same model runs in reverse - dialect speech in, transcribed text out.
upload audio · Egyptian
Drop an audio file here
WAV, MP3, or M4A
egyptian-voice-note.wavuploaded
transcribed text
الكلام المصري له إيقاع خاص ومميّز جدًا.
§ 07 - Film
See Natiq in action.
A short introduction to dialect-native Arabic speech, presented in Arabic.
§ 08 - API
Build with the API.
The same voices and dialects, inside your own product. One HTTP request returns audio, paid from the same credits as the app.
- POST /v1/speechText in, a WAV file out.
- POST /v1/speech/streamAudio starts arriving in about a second.
- POST /v1/transcribeArabic audio in, text out.
speech.sh
curl https://www.mithal.ai/api/v1/speech \
-H "Authorization: Bearer mth_live_..." \
-H "Content-Type: application/json" \
-d '{
"text": "مرحبا بك في منصة مثال",
"voice": "adam",
"dialect": "msa"
}' --output speech.wav200 OK · audio/wav · X-Credits-Charged: 21
§ 09 - Questions
Before you start.
What can I do with Mithal?
Turn text into speech in Egyptian, Saudi, Emirati or Modern Standard Arabic, clone a voice from a few seconds of audio, and transcribe Arabic recordings. Everything works in the app and through the API.
Is there an API?
Yes: speech, streaming speech and transcription. Once you have access, create a key on the API keys page. The API uses the same credits as the app.
How does pricing work?
You buy credits. Speech costs 1 credit per character of text, and transcription costs 2 credits per second of audio. Packs start at $19 for 100,000 credits. Purchased credits don't expire while your account is active, and a request that fails is refunded automatically.
How long can a clip be?
Up to 1,200 characters per generation in the app, which is about two minutes of speech. For longer scripts, generate them in parts.
Can I control emotion or speaking style?
Not yet. Today you choose the dialect, the voice (or a voice you cloned) and the generation settings. Emotion and style controls are on our roadmap.
Can I use the audio commercially?
Yes, for lawful purposes under our Terms of Service, including inside your own apps through the API. Reselling access to Mithal itself isn't allowed.
Can I clone anyone's voice?
Only with the speaker's explicit permission. That includes your own voice, and voices you have the rights to.
Is my audio used to train your models?
We don't use your private content to train general-purpose models without telling you first and getting any consent the law requires. Your history is kept in private storage in your account.
Why do I have to request access?
We're letting people in in batches while we scale. Request access with your email, and we'll send you a sign-in link once you're approved.
About mithal
Mithal is an Arabic language research project built by Ruliad, a UAE-based AI research lab.
Our first model is Natiq: an Arabic speech system developed around the language as it is actually spoken—across dialects, registers, and mixed-language use.