Arabic speech to text API: a developer's guide to transcribing Gulf Arabic

Most Arabic speech-to-text APIs will happily take your audio. The problem shows up in the text that comes back: a Gulf speaker says one thing, and the transcript reads like a newsreader said something else. This guide walks through transcribing Arabic with the nutq Arabic speech to text API — the request, the response, the cost — and the places where Arabic transcription usually fails, so you can test for them before you ship.

What does an Arabic speech to text API actually do?

It takes a recording and returns the words in it as text. For Arabic that simple job has three hard parts:

1. Dialect. People in Dubai, Riyadh and Amman do not speak Modern Standard Arabic (MSA) on the phone. Many models are trained mostly on MSA and "correct" dialect into it, which changes what was said. We cover this in detail in Gulf Arabic vs MSA transcription.
2. Audio quality. Much of the Arabic audio worth transcribing — support calls, sales calls, voice notes — is narrow-band and noisy. Accuracy measured on clean broadcast audio tells you little about a call centre.
3. Code-switching. Gulf speech mixes in English words mid-sentence. A model that forces one language per file mangles them.

How do you transcribe Arabic audio with one API call?

Create a key in the dashboard, then post the file:

curl -X POST https://nutq.dev/v1/transcribe \
  -H "Authorization: Bearer $NUTQ_API_KEY" \
  -F "file=@call.mp3"

There is no SDK to install. The same request from JavaScript:

const body = new FormData()
body.append('file', file)

const res = await fetch('https://nutq.dev/v1/transcribe', {
  method: 'POST',
  headers: { Authorization: `Bearer ${process.env.NUTQ_API_KEY}` },
  body,
})
const { text, usage } = await res.json()

And from Python:

import requests

with open("call.mp3", "rb") as f:
    res = requests.post(
        "https://nutq.dev/v1/transcribe",
        headers={"Authorization": f"Bearer {NUTQ_API_KEY}"},
        files={"file": f},
    )
print(res.json()["text"])

What comes back?

{
  "text": "هلا وغلا فيك، شحالك اليوم؟",
  "language": "ar",
  "language_source": "detected",
  "language_confidence": 1,
  "usage": { "audio_seconds": 13.23, "cost_usd": 0.00056 },
  "model": "nutq-listen-1"
}

Three fields are worth knowing about:

• language_source says how the language was chosen: requested (you named it), detected (you sent auto), corrected (you named one, but the recording was clearly another), or fallback (too short to judge, so Arabic).
• usage is what this request actually cost — not an estimate. Log it next to the transcript and your billing reconciles itself.
• model names the engine that served the request, so any transcript can be traced to a version.

Which Arabic dialects and languages are supported?

If you know the language, pass it as language=ar (or any code above). If the recording is clearly in another language, the API transcribes what was actually spoken and tells you so, rather than forcing the wrong language onto it.

How accurate is it on real-world audio?

We report one test, because it is the one that matches how Arabic audio is actually recorded. On a genuine narrow-band phone call-in, nutq made 40% fewer character errors than the leading Arabic vendor. The wider evaluation scored 986 clips — broadcast, spontaneous interviews and telephony — against human transcripts. The details, and why phone audio breaks most models, are in Transcribing Arabic phone calls.

Speed matters for batch work too: transcription runs about 22× faster than realtime, so an hour-long recording comes back in under three minutes.

What does Arabic transcription cost?

Files can be up to 100 MB and one hour long; longer recordings are split and stitched for you, so you send one request either way.

How do you test an Arabic speech API before committing?

Do not test on clean samples. Take twenty recordings that look like your production traffic — the real phone line, the real accents — and have a native speaker transcribe them by hand. Then compare each API's output against those transcripts, counting character errors. Look specifically at:

• dialect words turned into their MSA equivalents,
• English words inside Arabic sentences,
• names, numbers and product terms,
• the last seconds of each file, where some pipelines drop audio.

Your $5 of free credit covers far more audio than that test needs.

Can it run on our own servers?

Yes. For banks, government and anyone whose recordings cannot leave the building, the same models run inside your network on a single 24 GB GPU, with no audio sent to us and no per-minute fee. Talk to us about deployment.

Where to go next

• The full API reference covers streaming, errors and OAuth for apps that act on a user's behalf.
• Turning text back into speech? See Arabic text to speech with voice cloning.
• Captioning video? Auto Arabic subtitles shows the no-code route.