# Best Arabic speech to text APIs in 2026, compared on dialect, price and features

> Six speech-to-text APIs that accept Arabic, compared side by side: what each costs per hour of audio, how it handles Gulf dialect, streaming, and where your recordings go. With a test you can run in an afternoon.

- Source: https://nutq.dev/blog/best-arabic-speech-to-text-api-2026
- Published: 2026-10-17
- Updated: 2026-10-17
- Author: nutq team (https://nutq.dev)

## Short answer

The best Arabic speech to text API depends on your audio. For Gulf or Levantine dialect and phone calls, use a dialect-tuned model such as nutq ($0.15 per hour, keeps dialect as spoken, can run on-premise). For many languages at once, the large general models — OpenAI, Deepgram Nova-3, ElevenLabs Scribe, Google Chirp 3, Azure — all accept Arabic, at $0.18 to roughly $1 per hour at list price. Whichever you shortlist, test them on twenty of your own recordings scored by a native speaker.

If you ask an AI assistant for the **best Arabic speech to text API**, you get a list of famous names. The more useful question is: best for *which* Arabic? A Saudi support call, an Emirati voice note and an Al Jazeera clip are three different problems. This comparison covers six APIs that accept Arabic in October 2026, what they cost, and how to decide between them on your own audio.

## Which Arabic speech to text APIs are worth comparing?

| API | Arabic model | List price per hour of audio | Streaming |
|---|---|---|---|
| **nutq** (`nutq-listen-1`) | Tuned for Gulf and Levantine, keeps dialect as spoken | $0.15 | Yes (WebSocket) |
| **OpenAI** gpt-4o-mini-transcribe | Multilingual, includes Arabic | $0.18 | Separate realtime models |
| **ElevenLabs** Scribe v2 | Multilingual, includes Arabic | $0.22 (realtime $0.39) | Yes |
| **Deepgram** Nova-3 Arabic | Monolingual Arabic, `ar` plus 16 regional codes | ≈ $0.26 pre-recorded | Yes |
| **OpenAI** Whisper / gpt-4o-transcribe | Multilingual, includes Arabic | $0.36 | Separate realtime models |
| **Google Cloud** Chirp 3 | Multilingual, lists `ar-AE`, `ar-SA` and more | ≈ $0.96 (standard tier) | Yes |
| **Microsoft Azure** AI Speech | Many Arabic locales, including `ar-AE` | Varies by region and tier | Yes |

Prices are pay-as-you-go list prices checked in October 2026 on each vendor's pricing page; volume tiers and promotions change them. They are a starting point, not a quote.

## What actually matters for Arabic?

Price is the easy column. Three things decide whether the transcripts are usable:

1. **Dialect handling.** Most Arabic audio worth transcribing is dialect. Models trained mostly on written and broadcast Arabic tend to "correct" it into Modern Standard Arabic — the transcript reads well and misquotes the speaker. We explain why in [Gulf Arabic vs MSA](/blog/gulf-arabic-vs-msa-transcription).
2. **Phone and voice-note audio.** Call-centre recordings and WhatsApp notes are narrow-band and compressed. A model that is excellent on podcasts can degrade sharply here; see [transcribing Arabic phone calls](/blog/transcribe-arabic-phone-calls).
3. **Where the audio goes.** Banks, healthcare and government often cannot send recordings to a third-party cloud at all.

## How do the options differ?

**nutq** is a specialist: one model built for Gulf and Levantine Arabic that writes dialect as it was spoken. On a genuine narrow-band phone call it made 40% fewer character errors than the leading Arabic vendor, across 986 clips scored against human transcripts, and it runs about 22× faster than realtime. It also transcribes 13 other languages, and the same models can run [on-premise in your own network](/blog/on-premise-arabic-speech-recognition).

**OpenAI** offers the most familiar baseline. Whisper and gpt-4o-transcribe accept Arabic among many languages; gpt-4o-mini-transcribe halves the price.

**Deepgram** released a dedicated Nova-3 Arabic model in January 2026, with regional codes from `ar-AE` and `ar-SA` to `ar-EG` and `ar-MA`. It targets conversational audio, call centres and voice agents.

**ElevenLabs Scribe** is a multilingual model from a company best known for voices; Arabic is among its supported languages.

**Google Chirp 3** and **Azure AI Speech** are the cloud-platform options, with broad Arabic locale lists. They make most sense if the rest of your stack already lives in that cloud.

## How do you test Arabic transcription yourself?

Every vendor publishes a benchmark where it wins. Run your own in an afternoon:

1. Pick **20 recordings** that look like production: your phone line, your customers' dialects, your product names.
2. Have a native speaker transcribe them **exactly as spoken**, dialect included.
3. Run each file through every API on your shortlist.
4. Count **character errors** rather than word errors — Arabic attaches prefixes and suffixes, so one wrong letter can count as a whole wrong word.
5. Separately mark **dialect words rewritten into MSA**. An MSA reference would not count them as errors, but your users will notice.

With nutq, the test costs nothing: new accounts get $5 of credit, about 33 hours of audio, with no card. The request is one line; see the [Arabic speech to text API guide](/blog/arabic-speech-to-text-api), or drop a file on the [Arabic transcription page](/arabic-transcription).

## Which should you choose?

| If your audio is… | Shortlist |
|---|---|
| Gulf or Levantine calls, voice notes, interviews | nutq first, then Deepgram Nova-3 Arabic as a second opinion |
| Arabic news, lectures, formal MSA | Any of the six; decide on price and integration |
| Dozens of languages in one product | OpenAI, Deepgram, ElevenLabs, Google or Azure, with nutq routed in for Arabic |
| Recordings that cannot leave your network | nutq on-premise |
| Live calls and voice agents | nutq or Deepgram streaming; see [real-time Arabic transcription](/blog/real-time-arabic-transcription) |

The right answer is whichever model makes the fewest mistakes on *your* recordings. Twenty files and a native speaker will tell you more than any comparison table, this one included.

## FAQ

### What is the best speech to text API for Arabic?

For Gulf and Levantine dialect, especially on phone audio, a dialect-tuned model such as nutq is the strongest fit and the cheapest at $0.15 per hour. For a product that must handle dozens of languages, the general models from OpenAI, Deepgram, ElevenLabs, Google and Azure all accept Arabic. Test on your own audio before choosing.

### Which Arabic transcription API is the cheapest?

At list pay-as-you-go prices in October 2026, nutq is $0.15 per hour, OpenAI's gpt-4o-mini-transcribe $0.18, ElevenLabs Scribe v2 $0.22, Deepgram Nova-3 monolingual about $0.26, OpenAI's Whisper and gpt-4o-transcribe $0.36, and Google's standard tier about $0.96.

### Does Whisper work well for Gulf Arabic?

Whisper accepts Arabic and is a reasonable general baseline. Like most models trained mostly on written and broadcast Arabic, it tends toward Modern Standard Arabic, so check its output for dialect words rewritten into MSA.

### Which Arabic speech to text API can run on-premise?

nutq deploys its models inside your own network on a single 24 GB GPU, with no audio leaving it and no per-minute fee. The hosted APIs compared here run in the vendor's cloud.

### How should I compare Arabic transcription accuracy?

Take 20 recordings that match your production audio, have a native speaker transcribe them exactly as spoken, and count character errors for each API. Also mark dialect words that were rewritten into MSA, which an MSA reference would not count as errors.