Deepgram vs AssemblyAI for Arabic speech-to-text

At a glance

List prices in US dollars before tax and volume discounts. Sources: Deepgram pricing, Deepgram Arabic, AssemblyAI pricing and AssemblyAI models.

Which model handles Arabic

Deepgram’s language overview lists Arabic under Nova-3 as ar and sixteen regional codes, such as ar-EG and ar-MA; Arabic does not appear under Nova-2. Its Arabic page names Modern Standard, Egyptian, Levantine, Gulf, Maghrebi and Iraqi Arabic, for streaming and batch. Deepgram’s multilingual mode covers a separate list of languages that does not include Arabic, so Arabic should fall under the monolingual rate; its pricing page does not say so explicitly, so confirm it in your account.

AssemblyAI’s models page lists Arabic for Universal-2, which covers 99 languages, and for Universal-3.5 Pro, which covers 18. Its supported-languages page does not name Arabic dialects. AssemblyAI has also retired its older models, so check the model name in any example code you copy.

A dialect named on a product page is a statement of intent, not a measurement on your audio. Neither page tells you how a particular speaker, microphone or mix of Arabic and English will come out.

Streaming is where they differ most

Deepgram offers Arabic in real time on Nova-3, at a higher per-minute rate than batch, with speaker diarization charged extra when streaming.

AssemblyAI’s standard Universal-Streaming models list six languages and none of them is Arabic. Arabic real time goes through its Pro streaming model, priced at US$0.45 an hour. AssemblyAI’s own pages currently name that model inconsistently, as Universal-3.5 Pro Streaming on the models page and Universal-3.6 Pro Streaming in the streaming documentation, so confirm the model name in your account before you build on it.

Confirm each feature in Arabic

Neither company publishes a table of which features work for which language. Deepgram’s diarization documentation says it works in all available languages on Nova models, and not on its Whisper model. AssemblyAI says not every feature is available for every language: a request with an explicit language code is rejected if a feature is unavailable, while automatic language detection silently omits the feature.

So set language_code to ar explicitly on AssemblyAI, and on either service send one short Arabic file with speaker labels and word timestamps requested before you design around them. AssemblyAI can return SRT and VTT directly from its export endpoints; with Deepgram you build cues from the word timings yourself.

A third option: Kalemio

Kalemio is an Arabic-only API for recorded audio. A completed job returns JSON with the text and every word’s start and end time and speaker label, and an SRT file; the OpenAI-compatible endpoint can also answer as WebVTT. It does not offer real-time streaming. Each upload is limited to 20 MiB and 10 minutes per file, so longer recordings are split and submitted as separate jobs, and a paid account can have two jobs queued or running at once. Results are kept for 24 hours, so save them. The API guide documents each of these.

Its pricing is monthly rather than per minute. Starter is US$5 a month for 1,250 minutes, and Studio is US$13.93 a month for 5,000 minutes; included minutes reset each month and do not roll over. 100 hours a month would be Studio plus 2 top-ups of 500 minutes at US$5 each, US$23.93 in total. At low volume a monthly plan costs more per processed minute than a per-minute API. See Kalemio’s plans.

Kalemio’s published benchmark covers one narrow Levantine sample and does not include Deepgram Nova-3 or AssemblyAI’s current models, so it says nothing about how they compare. Read the benchmark method for its limits.

How to choose

• Deepgram if you need Arabic in real time, regional Arabic codes, or speaker labels included in the base price, once you have confirmed the rate Arabic is billed at.
• AssemblyAI if you transcribe recorded Arabic at volume and its list price per hour is lower for you, or you want SRT and VTT straight from the API, and you have confirmed with a test that the features you need work in Arabic. Its US$17 total includes speaker labels, which are not confirmed for Arabic.
• Kalemio if your product is Arabic-only, works from recordings, and needs word timings, speaker labels and SRT in one result.

Whichever you shortlist, run the same Arabic recordings through each: separate dialects, phone and studio audio, and Arabic mixed with English. Have a fluent reviewer write the reference text, count wrong names, numbers and negations separately, and check timings and speaker changes by listening.

Review Kalemio’s monthly plans Read the integration guide