I cloned my voice in thirty seconds. The hard part was making it say the right word.

Why Arabic speech sounded robotic, and what a clone actually changes

Every Arabic text to speech model is trained overwhelmingly on Modern Standard Arabic, because that is what Arabic speech data mostly is. News, audiobooks, recitation, formal broadcast. So the model learns one way for Arabic to sound, and it applies that way to whatever you give it. Hand it a line of Levantine, Egyptian or Gulf dialect written in ordinary spelling and it reads the dialect words with Standard pronunciation. The result is understandable and nobody believes a person said it. Voice cloning changes where the sound comes from. Instead of generating speech from its own idea of Arabic, the model conditions on your reference recording and copies its accent and delivery. That is why a clone can speak your dialect when a built-in voice cannot, and it is also why the quality of the reference recording matters more than any setting you can adjust afterwards. It also explains the one failure in this guide that cannot be fixed. A local open-source model with a Standard-Arabic text frontend will still convert letters to sounds the Standard way, whatever voice you clone on top. Eight attempts across rewriting the text and changing the settings did not move it, because the settings were never the problem.

Three tools against one dialect line, and what free actually includes

Google released gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts on 2026-09-23. Replication needs a thirty second sample plus a verbal consent recording from the voice owner, which the system checks against the reference speaker, and the whole thing is available on the free tier. Every clip carries a SynthID watermark and replicated voices also carry C2PA content credentials, per Google's announcement. On the dialect line it was the best of the three I tried. ElevenLabs v4 lists Arabic in its supported languages, and it is what I used for the Syrian voice-over of a Reel I published on 2026-10-11. It works. The catch is lower down the pricing page: the free plan gives ten thousand credits a month, roughly ten minutes of speech, carries no commercial licence, and does not include voice cloning in any form. Instant cloning starts at six dollars a month on Starter and professional cloning at twenty two dollars on Creator. Chatterbox Multilingual is open source under MIT, runs on your own machine, costs nothing and watermarks nothing. I ran it on a MacBook Pro M1 Pro. English is good and Modern Standard is good. Syrian dialect is bad, and it stayed bad across eight attempts that covered both rewriting the text and changing the generation settings. It is also slow: about six seconds of audio took about seventy five seconds to produce. The money is not the deciding factor. Gemini bills audio at twenty five tokens per second, and nine dollars per million output tokens until the end of 2026, which puts a forty five second voice-over at about one cent. ElevenLabs charges one credit per character, which on the Creator plan works out near thirteen cents for the same clip. Both are cheap. Choose on whether the free tier lets you hear your own voice before you pay, and on whether a watermark matters to you.

The wrong-word failure, and the rule that stops it

Arabic script leaves out short vowels and gemination, so a model reading undiacritized text has to guess them. In most languages a bad guess sounds wrong. In Arabic it frequently lands on a different word that exists, means something else, and fits the sentence. On 2026-10-03 it happened to me twice. One line should have said "he wrote it" and the clip said "he made someone else write it". The other is the one above, where reaching became a lift in a car. Both clips sounded natural, both were grammatical, and neither is visible if you reread your script instead of listening to the file. The rule that fixes most of this is to diacritize every letter, not the hard words. The only letters that stay bare are the ones that cannot carry a mark anyway: long alef, alef maqsura, madda alef, waw after damma and yaa after kasra because they are long vowels, the lam of the article before a sun letter that already carries shadda, and a word-final ta marbuta. I do not send any Arabic line to generation until an automatic check confirms there is no bare letter left. Then there is the case that survives the rule. The next day I wrote a word fully diacritized, meaning "it reaches you", and it still came out as "it drives you somewhere". Some verbs have a doubled form that is itself a common verb, and the model leans toward it even when the marks say otherwise. More diacritics will not help. A different word will: I swapped it for an unambiguous synonym and the problem disappeared. These are the roots I now treat as unsafe unless I can listen to the output myself: wasala, kataba, alima, arafa, darasa, fahima, khalasa, waqafa, samia, rajaa, talaa, nazala, dakhala, qaruba, kasara, ittafaqa. Whether Gemini needs the same discipline, I do not know yet. Google's documentation says nothing about diacritics. The honest test is to generate the same line twice, once marked and once bare, and listen to both before you build a workflow on the answer.

Record the reference properly, because it is the only lever you get

The Voices API reference has no endpoint that retrains or edits an existing voice. There is create, list, get and delete. A remix is a create call that takes an existing voice as a base plus a prompt describing the change, and it returns a new voice rather than improving the old one. So the reference recording decides the quality of everything that follows. Record one continuous take of twenty five to thirty five seconds in a quiet room, with no fan, no air conditioning and no open window. Keep the microphone fifteen to twenty centimetres away and slightly off to the side so your plosives miss it. Write yourself a short script in your dialect that contains a statement, a question and a call to action, so the model hears your intonation range instead of one flat contour, and put one English word in it if you code-switch in normal speech. Do not edit or denoise the take. The model wants your room as it is. Speak at the energy you want back. This is the part I got wrong first: the clone copies delivery, so a tired reference produces a tired voice. Record the consent clip separately, reading the exact phrase the interface shows you, word for word, because any deviation fails validation. Two more things worth knowing before you design around the API. Output comes back as WAV at 24 kHz, mono, 16-bit, per the speech generation docs, and a single request supports two speakers only when both are ready-made voices. A dialogue in your own cloned voice has to be generated turn by turn and joined afterwards.

استنسخ صوتك بلهجتك مجانًا، ثم احذر كلمة تُنطق بمعنى آخر

سجّلت عينة صوتية مدتها ⁨30⁩ ثانية فور استيقاظي، وكان صوتي فيها متعبًا ولم تكن أفضل تسجيلاتي. وضعتها في ⁨Google AI Studio⁩ فخرج صوت يتكلم بلهجتي السورية، والنتيجة كانت ممتازة.

فاستنساخ الصوت بلهجتك صار شيئًا تجربه مجانًا قبل أن تدفع قرشًا. وإن سبق أن جرّبت النطق الآلي بالعربية وسمعت نشرة أخبار، فالسبب أن اللهجة كانت تأتي من تدريب الصوت. في نماذج ⁨Google⁩ الجديدة تأتي من تسجيل قصير تقدمه أنت، ولهذا صارت لهجتك ممكنة أصلًا.

جرّبت ثلاث أدوات. ⁨Gemini⁩ يعطيك الاستنساخ في طبقته المجانية، و⁨ElevenLabs⁩ لا يعطيه في خطته المجانية أصلًا، و⁨Chatterbox⁩ مجاني ويعمل على جهازك لكنه لم يتقن الشامية معي مهما غيّرت، وعلى ⁨MacBook Pro M1 Pro⁩ استغرق توليد ست ثوانٍ من الصوت أكثر من دقيقة. في الصفحة الأسعار الدقيقة لكل واحدة، وما يعنيه مجاني في كل حالة.

أما ما لا يحذّرك منه أحد فهو النطق. الكلمة العربية بلا تشكيل تحتمل أكثر من قراءة، والنموذج يخمّن، وإذا أخطأ خرجت كلمة سليمة بمعنى آخر، ويفهم المستمع جملة غير جملتك دون أن يشعر بأي خطأ. حدث لي ذلك مع ثلاث كلمات، وفي إحداها كان التشكيل كاملًا ومع ذلك أخطأ النموذج.

إن كنت تصنع ريلز أو فيديوهات قصيرة بالعربية فالصفحة تختصر عليك تجارب كثيرة: ماذا تجرّب، وبأي ترتيب، وماذا تفحص قبل أن يسمعه جمهورك. فيها قاعدة التشكيل التي أطبقها منذ ⁨2026-10-03⁩، وقائمة جذور أستبدل كلماتها حين لا أستطيع الاستماع بنفسي، وسؤال تركته الوثائق دون جواب عن التشكيل مع ⁨Gemini⁩ مع طريقة لتختبره بأذنك. وفي آخر الصفحة مربع نسخ فيه خطوات التسجيل والنشر جاهزة. اقرأها قبل أن تنشر أول مقطع بصوتك.

تابعني على ⁨Instagram⁩:

التسجيل في الدورة: https://halaqa.app/enrollment?course=start-with-ai

الخطوات

Prompt

# ARABIC VOICE CLONE: RECORD, GENERATE, CHECK
# Use before you clone a voice and before you publish anything it generated.

# ─── 1. THE REFERENCE RECORDING (do this once, properly) ───
# Room:   quiet, no fan, no AC, no open window. Soft surfaces beat a bare room.
# Mic:    15-20 cm away, angled slightly off-axis so plosives miss it.
# Take:   ONE continuous take, 25-35 seconds. No edits, no denoise, no EQ.
# Energy: speak at the level you want in the output. A tired take clones tired.
# Script: your dialect, and it must contain all of:
#           - one statement
#           - one question
#           - one call to action
#           - one English word, if you code-switch in normal speech

# ─── 2. THE CONSENT RECORDING (separate file) ───
# Read the exact phrase the interface shows you, word for word.
# Any deviation fails validation. Your own voice = you record it yourself.
# Someone else's voice = you need the recording FROM THEM first.

# ─── 3. BEFORE YOU GENERATE: THE ARABIC TEXT ───
# [ ] Every Arabic letter carries a diacritic.
#     Exempt (cannot carry one): long alef, alef maqsura, madda alef,
#     waw after damma, yaa after kasra, article lam before a shadda'd sun
#     letter, word-final ta marbuta.
# [ ] No word from the ambiguous-gemination roots below, OR you will listen
#     to the output yourself before using it:
#       wasala  kataba  alima   arafa   darasa  fahima  khalasa  waqafa
#       samia   rajaa   talaa   nazala  dakhala qaruba  kasara   ittafaqa
#     Fix: swap in an unambiguous synonym. More diacritics will not help.

# ─── 4. AFTER YOU GENERATE ───
# [ ] Listen to the WHOLE clip. Do not read the script and assume.
#     The failure mode is a real word with the wrong meaning, in your own
#     voice, in a grammatical sentence. Only your ear catches it.
# [ ] Dialogue in a cloned voice: generate each turn separately and join.
#     Two speakers in one request works only with ready-made voices.

# ─── 5. KNOW WHAT YOU ARE SHIPPING ───
# Cloned output from Google's models carries a SynthID watermark and C2PA
# content credentials. Inaudible, present in the file. Fine for your own
# channels. Decide deliberately before a client deliverable.

# ─── THE TEST WORTH RUNNING ONCE ───
# Generate the same line twice: fully diacritized, and bare.
# Listen to both. That tells you whether your tool needs the discipline
# in section 3, and no documentation will answer it for you.