# Arabic speech to text API in 2026: how to test Gulf, Egyptian, Levantine and Maghrebi transcription > Arabic speech to text is hard because people speak dialects, mix in English and record on noisy phones. This guide shows how to test any Arabic transcription API on your own audio, and what the BroadNet Inference API offers for Gulf, Egyptian, Levantine and Maghrebi Arabic. - Canonical URL: https://www.broadnet.me/blog/arabic-speech-to-text-api-dialects-2026 - Language: en - Last updated: 2026-11-08 - Publisher: BroadNet (https://www.broadnet.me) - Published: 2026-11-08 - Author: BroadNet Editorial Team An **Arabic speech to text API** turns recorded Arabic speech into Arabic-script text that your software can search, summarize and store. The hard part is dialect: Gulf, Egyptian, Levantine and Maghrebi speech differ from Modern Standard Arabic, and speakers mix in English. Choose by testing word error rate on your own audio. The [BroadNet Inference API](https://www.broadnet.ai/inference?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) transcribes Arabic in these dialects for $0.0042 per audio minute, without storing your audio. > Facts checked at their sources on 8 November 2026. Research facts come from the Open Universal Arabic ASR Leaderboard and its paper, a bilingual Arabic-English leaderboard, and the ArzEn, Casablanca, MASC, SADA and FLEURS dataset pages, each linked inline. Product facts come from broadnet.ai/inference, its pricing, model, privacy and developer pages, and its llms.txt. Prices and rankings change, so check the linked page before you buy. This table is the short version of the guide. Use it as a checklist when you compare any Arabic transcription service. | What to test | Why it matters for Arabic | How to test it | What the BroadNet Inference API documents | |---|---|---|---| | Dialect accuracy | Error rates change a lot between dialects and recording conditions | WER and CER on 1-3 hours of your own audio, per dialect | Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic | | Script and code-switching | Speakers mix English words; some systems switch to Latin letters | Count words written in the wrong script | Arabic model stays in Arabic script; always send language=ar | | Numbers, names, codes | One wrong digit breaks a phone number, an amount or an order | Check every number and name by hand in your test set | A fine-tune for numbers, dates, names and codes, on request | | Speakers and timestamps | Call QA needs to know who spoke and when | Check the response format for speaker labels and times | Not documented: output is Arabic text, as JSON or plain text | | Speed | Batch jobs and voice agents need different speeds | Time a fixed set of files end to end | 45x realtime per stream and 190x batched, measured by BroadNet | | Data handling and region | Contact-centre audio is often personal data | Read the retention statement and ask where processing happens | Zero data retention; US GPUs; MENA or Europe on request | | Price | Hours of audio add up fast | Price one month of your real volume | $0.0042 per minute (about $0.25 per hour); $5 free credit | ## Why is Arabic speech recognition harder than English? Arabic speech recognition is harder because one written standard sits on top of many spoken dialects, and written Arabic usually leaves out short vowels. A system must hear dialect sounds and words, then write them in a spelling that has no single standard. The authors of the [Open Universal Arabic ASR Leaderboard paper](https://arxiv.org/abs/2412.13788) name three causes: rich morphology, dialect variation and the lack of diacritics in written text. They also note that Modern Standard Arabic is used in formal settings, while regional dialects differ in both pronunciation and vocabulary. In practice, this creates four problems for a transcription system: - **Dialect words.** A Gulf, Egyptian or Moroccan speaker uses everyday words that never appear in news Arabic. - **Spelling choices.** The same dialect word can be written in several ways. - **Code-switching.** Speakers move between Arabic and English, or Arabic and French, inside one sentence. - **Noisy audio.** Calls, street recordings and voice notes are much harder than studio speech. The numbers show the gap. On the [Open Universal Arabic ASR Leaderboard](https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard), checked on 8 November 2026, the best system scored about 5% WER on the Common Voice test set. The best score on [SADA](https://www.kaggle.com/datasets/sdaiancai/sada2022), a Saudi test set with 10 dialects, was about 29%. Across all 39 systems, SADA scores ran up to 89%. ## Which Arabic dialects does your audio contain? Start by listing the dialects your customers actually speak, then make sure your test set covers each one. Most buyers need some mix of these groups: - **Gulf (Khaleeji):** Saudi Arabia, the UAE, Qatar, Kuwait, Bahrain and Oman, with strong regional variants inside Saudi Arabia. - **Egyptian:** Egypt, and widely heard in regional film and TV. - **Levantine:** Lebanon, Jordan and Palestine. - **Iraqi:** a distinct group that some services do not list separately. - **Maghrebi:** Morocco, Algeria and Tunisia, with many French loanwords. - **Modern Standard Arabic:** news, speeches, formal announcements and read text. The [Casablanca dataset](https://arxiv.org/abs/2410.04527) collects transcribed speech in 8 dialects, including Emirati, Egyptian, Jordanian, Moroccan and Algerian, with labels for code-switching. The [ArzEn corpus](https://aclanthology.org/2020.lrec-1.523) contains 12 hours of spontaneous Egyptian Arabic-English code-switched speech from 38 bilingual speakers. Code-switching remains hard even for strong systems. On a public [bilingual Arabic-English leaderboard](https://huggingface.co/spaces/QCRI/AURA-Arabic-English-ASR-leaderboard), the leading systems scored about 30% WER on ArzEn's mixed Egyptian Arabic and English speech. The same systems scored about 6% on read Arabic speech from [FLEURS](https://arxiv.org/abs/2205.12446). If your callers mix languages, include that speech in your test. ## How do you measure Arabic transcription accuracy? Measure word error rate (WER) and character error rate (CER) on a sample of your own recordings, with a human-checked reference transcript. Public leaderboards only help you build a shortlist. WER counts the words a system substitutes, deletes or inserts, divided by the number of words in the reference. CER does the same at the character level. CER is useful for Arabic because a small spelling difference, such as a hamza, counts as a whole wrong word in WER but only one wrong character in CER. Follow these steps: 1. Collect 1 to 3 hours of real audio per dialect: calls, voice notes or media, with the same quality you will send in production. 2. Have native speakers write a reference transcript in Arabic script. Agree on spelling rules before they start. 3. Normalize both texts the same way before scoring. The [leaderboard paper](https://arxiv.org/abs/2412.13788) removes punctuation and diacritics, normalizes hamza and madda forms, and converts Eastern Arabic numerals to Western numerals. 4. Score each system on the same files with the same settings, and report WER and CER per dialect, not only one average. 5. Check numbers, names and product codes by hand. A low WER can hide a wrong digit that matters more than ten small errors. 6. Repeat the test every few months. New models appear often, and leaderboard ranks change. Always note the conditions behind a published figure. The leaderboard uses multi-dialect test sets such as [MASC](https://ieee-dataport.org/open-access/masc-massive-arabic-speech-corpus), which holds about 1,000 hours of speech from more than 700 YouTube channels, and SADA. A score on clean read speech says little about a noisy call. ## What else should you check besides accuracy? Check what the output contains, how fast it arrives, which file formats are accepted and where the audio is processed. These details decide whether a transcript fits your workflow. - **Speaker labels (diarization).** Call quality reviews need to know who said what. If a service does not label speakers, record calls in stereo with the agent and the customer on separate channels, then transcribe each channel. - **Timestamps.** Media subtitles and audio search need word or segment times. Check the response format before you build. - **Batch or live.** Overnight call analysis needs throughput. A live voice agent needs fast answers. - **File formats and size.** Check the accepted formats and the maximum file size, and plan to split long recordings. - **Data retention.** Ask whether audio or transcripts are stored, logged or used for training, and get the answer in writing. - **Processing region.** Many banks, governments and health providers must keep data in a set country or region. Ask your legal team which rules apply. ## Where does Arabic speech to text pay off? Arabic transcription pays off wherever people already speak more than they type: contact centres, media archives and voice messages. In each case the transcript becomes data you can search, check and feed into other tools. - **Contact-centre QA.** Transcribe every call, not a small sample. Search for complaint words and check that agents read required statements. - **Media and archives.** Make interviews, podcasts and broadcasts searchable. - **WhatsApp voice notes.** Customers often send voice notes instead of typing. Transcribing them lets a support team or an AI agent read, route and answer them. - **Voice agents.** An AI agent can only act correctly if it hears the phone number, amount or date correctly. If you are planning AI agents on top of transcripts, read our guides to an [enterprise agentic AI platform](https://www.broadnet.me/blog/enterprise-agentic-ai-platform-2026) and to [turnkey AI agents for hospitality, retail and healthcare](https://www.broadnet.me/blog/turnkey-ai-agents-hospitality-retail-healthcare-2026). ## How does the BroadNet Inference API transcribe Arabic? The [BroadNet Inference API](https://www.broadnet.ai/inference?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) is an API that runs speech, document and chat models on GPUs that BroadNet operates. For Arabic audio, you send a file to the transcription endpoint with the Arabic model and language=ar, and you get Arabic-script text back. It uses the common chat-completions and audio-transcriptions request format, so existing client libraries work once you change the base URL. These facts come from BroadNet.AI's own pages. - **Dialects.** The Arabic model covers Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic, according to the [Inference API page](https://www.broadnet.ai/inference?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026). - **Arabic script.** The site says the Arabic model stays in Arabic script, while the multilingual model can drift into English on mixed speech. It recommends the Arabic model for Arabic. - **Model.** The Arabic model is an open-weight model with 2 billion parameters under the Apache-2.0 licence, served on BroadNet hardware. - **Measured accuracy.** In BroadNet's own benchmark, the model scored 21.1 WER on real dialect speech (MASC), 23.4 on a Gulf, Egyptian and Levantine set, and 3.8 on clean read speech (FLEURS Arabic). - **Public ranking.** On 8 November 2026 the [Open Universal Arabic ASR Leaderboard](https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard) listed the same base model 4th of 39 systems, with an average WER of 24.19%. - **Speed.** BroadNet reports 45x realtime per stream and about 190x realtime when you send requests in parallel. - **Fine-tune on request.** A BroadNet fine-tune trained on real call audio targets phone numbers, amounts, dates, order codes and names over telephone lines and in noisy voice notes. It is issued as a private model ID and can be tuned further on your own recordings. - **Chat and documents.** The same API serves vision-language models, including Gemma 4 VL 26B, that read Arabic and English documents and return JSON fields. On data, the [data-handling statement](https://www.broadnet.ai/inference-privacy?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) says prompts, audio, images and transcripts are held in memory for one request and then discarded. Content is not logged or used for training, and an automated audit checks this configuration every hour. BroadNet keeps only content-free records such as timings, token counts and cost. Processing happens on BroadNet-operated hardware in the United States, and regional processing in MENA or Europe is available on request. For larger or regulated workloads, a private endpoint gives you a reserved GPU or node, your chosen model (including the Arabic fine-tunes), your own key and limits, and the region you need. ## What does Arabic transcription cost on BroadNet? The public Arabic model costs $0.0042 per minute of audio, which is about $0.25 per hour. New accounts get $5 of free credit, valid for 30 days with no card needed, which the [pricing page](https://www.broadnet.ai/inference/pricing?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) says covers about 19 hours of Arabic speech to text. | Option | Price (USD) | Availability | |---|---|---| | Arabic transcription model | $0.0042 per audio minute | Live, public model ID | | Arabic model, BroadNet fine-tune for calls | $0.005 per audio minute | On request, private model ID | | Multilingual transcription (14 languages) | $0.002 per audio minute | Live, public model ID | | Gemma 4 VL 26B (chat, images, documents) | $0.10 input / $0.22 output per 1M tokens | Live, public model ID | | Hourly key (Gemma 4 only) | $0.50 per one-hour key | Live, minted by API | | Private endpoint | Quoted to your traffic | On request | For example, 10,000 minutes of Arabic calls a month would cost $42 on the public Arabic model at the published rate. Volume and reserved-capacity pricing are available on request. ## What the BroadNet Inference API does not do The API is a file-based transcription and model-serving service, and some features that buyers expect are not documented. - **No documented speaker labels or timestamps.** The model pages describe the output as Arabic-script text in JSON or plain text. - **No documented live streaming transcription.** You upload one audio file per request. Streaming is documented for chat, not audio. - **Dialect list.** Iraqi Arabic and other regional varieties are not named separately on the site. Test them with your own audio. - **Formats and size.** The developer reference documents WAV and MP3 with a maximum file size of 26,214,400 bytes (25 MiB). The model page also lists M4A, OGG and FLAC, but the reference says other containers are not documented for the direct API yet. Convert WhatsApp voice notes to WAV or MP3 to be safe. - **Default region.** Standard processing is in the United States. Regional processing needs a request. - **Limits.** A pay-as-you-go key allows 30 requests per minute and 4 requests at once, and an account can hold up to 10 live keys. - **Fine-tune access.** The call-tuned Arabic model is not self-serve. You ask for it with a sample of your audio. ## How do you get started? You can hear the Arabic model on your own voice before you create an account. 1. Open the [voice playground](https://www.broadnet.ai/inference/playground?mode=voice&utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026). Record up to one minute of speech or upload a voice note. No sign-up or card is needed. 2. Create an account at [inference.broadnet.ai](https://inference.broadnet.ai/?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) with Google or an emailed sign-in link. The $5 credit is added with no card. 3. Create an API key in your account. Add your server IP addresses so the key is refused from anywhere else. 4. Follow the [developer reference](https://inference.broadnet.ai/docs?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026): set the base URL to https://inference.broadnet.ai/v1 and use any client library that supports the common chat-completions format. 5. Send each file to /v1/audio/transcriptions with the Arabic model ID and language=ar. Without language=ar, the model may write Arabic in Latin letters. 6. Run your WER and CER test per dialect, then ask for the call-tuned model or a private endpoint if you need them. BroadNet.AI is one of the [BroadNet Group brands](https://www.broadnet.me/brands). For agent deployments on top of transcription, see [BroadNet AI](https://www.broadnet.me/platforms/broadnet-ai) and BroadNet.AI's notes on [Arabic dialect variation in 14 markets](https://www.broadnet.ai/blog/arabic-dialect-handling?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026). **Try it on your own audio today.** Open the [BroadNet Inference API](https://www.broadnet.ai/inference?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026), record a voice note in your customers' dialect, and read the transcript in seconds. Then use the $5 free credit to test about 19 hours of your real calls. Visit [BroadNet.AI](https://www.broadnet.ai/?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) to compare the full platform. ## Frequently asked questions ### What is the best Arabic speech to text API? The best Arabic speech to text API is the one with the lowest error rate on your own audio, in your dialects, at a price you can sustain. Public leaderboards change often, so use them only to build a shortlist. Then test 2 or 3 services on the same recordings. The [BroadNet Inference API](https://www.broadnet.ai/inference?utm_source=broadnet.me&utm_medium=blog&utm_campaign=arabic-speech-to-text-api-dialects-2026) offers a free browser demo and $5 of credit for this kind of test. ### Can speech recognition understand Arabic dialects? Yes, modern speech recognition models can transcribe major Arabic dialects, but accuracy varies a lot by dialect and recording quality. Public benchmarks show much higher error rates on Saudi dialect speech than on read speech. Gulf, Egyptian, Levantine and Maghrebi Arabic are now commonly supported. Smaller regional varieties may need testing or a fine-tuned model. ### How accurate is Arabic speech to text? Accuracy depends on the audio. On clean read speech, strong systems reach single-digit word error rates. On noisy calls and dialect conversation, error rates of 20% to 30% or more are common, even for leading systems. Measure WER and CER on your own recordings, and check numbers and names by hand. ### How do I transcribe WhatsApp voice notes in Arabic? Export or receive the voice note file, convert it to WAV or MP3, and send it to an Arabic transcription API with the language set to Arabic. Use a model that keeps Arabic script, so the text works for search and routing. On the BroadNet Inference API, each file can be up to 25 MiB, and audio is not stored after the request. ### Does Arabic speech to text handle Arabic mixed with English? Many systems can, but mixed Arabic and English speech is still one of the hardest cases. On a public bilingual leaderboard, leading systems scored about 30% WER on code-switched Egyptian Arabic and English. Include mixed speech in your test, and check that Arabic words stay in Arabic script. ### Is my audio stored when I use the BroadNet Inference API? No. BroadNet.AI states that audio, prompts and transcripts are held in memory for one request and then discarded. Content is not logged or used for training, and an automated audit checks this every hour. BroadNet keeps only content-free records such as timings, token counts and cost. Processing is in the United States, with MENA or Europe on request. ### How much does Arabic speech to text cost on BroadNet? The public Arabic transcription model costs $0.0042 per minute of audio, about $0.25 per hour. A fine-tune trained on call audio costs $0.005 per minute and is available on request. New accounts get $5 of free credit, valid for 30 days with no card, which covers about 19 hours of Arabic audio.