Arabic text to speech with voice cloning: a Gulf Arabic TTS guide

Most Arabic text to speech sounds like the evening news: correct, formal, and nothing like the way anyone in the Gulf talks. That is fine for a train announcement. It is wrong for a customer-service agent, a brand voice, or a video voiceover aimed at people in Dubai or Riyadh. This guide covers Arabic text to speech with voice cloning — generating Gulf Arabic speech in a specific voice from a few seconds of audio.

What is Arabic voice cloning?

Voice cloning makes a text-to-speech model speak in a particular person's voice using a short sample of them talking, instead of a fixed stock voice. You give it three things:

1. reference — a recording of the voice, ideally six to eight seconds.
2. ref_text — exactly what is said in that recording. The model aligns the audio against it.
3. text — what you want the voice to say.

The result is a WAV file of your text, in that voice, with a Gulf accent. The model, nutq-speak-1, is trained on Emirati and Saudi speech, so it sounds Gulf rather than like broadcast MSA.

How do you generate Arabic speech with the API?

curl -X POST https://nutq.dev/v1/speech \
  -H "Authorization: Bearer $NUTQ_API_KEY" \
  -F "text=هلا وغلا فيك، شحالك اليوم؟" \
  -F "reference=@voice.wav" \
  -F "ref_text=نص التسجيل المرجعي" \
  --output speech.wav

The body of the response is the audio itself, so usage travels in the headers:

HTTP/1.1 200 OK
Content-Type: audio/wav
X-Nutq-Characters: 57
X-Nutq-Audio-Seconds: 5.672
X-Nutq-Cost-USD: 0.000285
X-Nutq-Model: nutq-speak-1

The same call in Python:

import requests

res = requests.post(
    "https://nutq.dev/v1/speech",
    headers={"Authorization": f"Bearer {NUTQ_API_KEY}"},
    data={"text": "هلا وغلا فيك", "ref_text": "نص التسجيل المرجعي"},
    files={"reference": open("voice.wav", "rb")},
)
open("speech.wav", "wb").write(res.content)

What makes a good reference recording?

The reference clip decides most of the quality. A few rules that matter:

Longer clips are accepted, but they are trimmed to about eight seconds and re-transcribed, which costs a round trip and can pick a less representative stretch. Trim it yourself.

Clone only voices you have permission to use — your own, or a speaker who has agreed to it.

How do you fix a word that is pronounced wrong?

Arabic is normally written without short vowels and read from context, so a word with two possible readings is a guess. تعد can be تَعُدّ or تُعَدّ; درس can be دَرَسَ or دَرَّسَ. The fix is to add diacritics to that one word:

تُعَدُّ دولة الإمارات من أبرز الدول

Mark only what is ambiguous — the way Arabic is normally written. Marking every word changes the delivery, not just the vowels: speech runs about half again as long and sounds like a recitation. We also tested automatic diacritizers and do not recommend them; they introduce errors of their own, especially in dialect.

How do you trade quality against speed?

For a live voice agent, lower steps are worth it: in a phone conversation the wait is more audible than the difference in quality. For a voiceover you will publish, use 64.

What does Arabic TTS cost?

$0.05 per 1,000 characters. A 1,000-character script — roughly a minute and a half of speech — costs five cents. New accounts get $5 of free credit with no card at sign-up, and failed requests are never billed.

Can it speak other languages?

Yes, with trained voices instead of cloning. Pass language as en, en-gb, es, fr, hi, it, ja, pt or zh, and a voice name such as af_heart or bm_george. reference and ref_text are ignored for these. GET /v1/voices lists what is available.

Where to use it

• Voice agents that answer customers in Gulf Arabic — the API reference has a ready Vapi configuration for both listening and speaking.
• Video voiceover and dubbing — see auto Arabic subtitles and dubbing for video.
• Transcribe first, then speak: pair it with the Arabic speech to text API to build a full voice loop.