Deepgram

Nova-3 covers Arabic, with seventeen language codes

The documentation lists Arabic among the languages of Nova-3, Deepgram’s strongest general transcription model, under the generic code ar and regional codes for the UAE, Saudi Arabia, Qatar, Kuwait, Syria, Lebanon, Palestine, Jordan, Egypt, Sudan, Chad, Morocco, Algeria, Tunisia and Iraq, plus one for Iran. Arabic is also part of Nova-3’s multilingual mode, which helps when speakers switch between Arabic and English mid-sentence, as they often do on Gulf customer calls.

A country code does not necessarily mean a model trained on that dialect; the docs list the codes without explaining how they differ. Arabic is also missing from Flux, the models built for real-time conversation with turn detection: Flux English is English only, and Flux Multilingual covers ten languages, none of them Arabic.

It hears Arabic but cannot speak it

Deepgram’s text-to-speech models, Aura-2 and Flux TTS (launched in August 2026), have no Arabic voice. Aura-2 covers English, Spanish, German, French, Dutch, Italian and Japanese, and Flux TTS launched in English only. An Arabic voice agent therefore means pairing Deepgram’s transcription with a voice from another provider such as ElevenLabs; the Voice Agent API has a cheaper tier for bringing your own text-to-speech.

Pricing: $200 of free credit, then per minute

Pay As You Go starts with $200 of free credit, no minimum and no credit card. Rates from the pricing page in October 2026:

Some add-ons cost extra, such as PII redaction at $0.0020/min and keyterm prompting at $0.0013/min. At those rates an hour of pre-recorded audio on Nova-3 Monolingual comes to about 26 cents before add-ons. The Growth plan starts at $4,000 a year, prepaid, for discounts of up to 20%.

Where your audio is processed

Deepgram offers an EU endpoint for processing inside the European Union, has announced India and Australia endpoints on its blog, and documents self-hosted deployment for enterprises. We found no Gulf or Middle East endpoint. If your country’s rules require audio to stay in-country, self-hosting is the route, and that means a conversation with sales.

Test it on your own dialect before writing code

The free credit is enough for a proper test. Take ten minutes of real audio you actually receive, not read-aloud Modern Standard Arabic, and transcribe it twice: once with the generic ar code and once with your country’s code. Compare the errors on numbers, names and colloquial words. If your callers mix Arabic and English, try the multilingual mode as well. For a wider comparison of Arabic transcription tools see our practical guide to transcribing Arabic audio; the closest direct alternative is AssemblyAI.

Official site: deepgram.com