To choose an Arabic speech-to-text API, transcribe 30 to 50 of your own clips per dialect with each candidate, compare word error rate per dialect, then weigh price per audio hour, file limits, streaming needs and data retention. A model that looks great on a benchmark can still fail on your call-centre audio.
Arabic is hard for speech recognition: one written standard sits on top of many spoken dialects, speakers switch to English mid-sentence, and most real audio is phone calls and voice notes.
Why does word error rate on your own audio matter most?
Word error rate (WER) is the share of words the model gets wrong: substituted, missed or invented. Lower is better. Public numbers are a starting point:
| Test set | What it is | BroadNet Transcribe Arabic | ElevenLabs Scribe v2 | Whisper large-v3 |
|---|---|---|---|---|
| FLEURS-ar | Clean read sentences | 3.8 | not published | not published |
| MASC-ar | Mixed real-world Arabic speech | 21.1 | 22.3 | 31.5 |
| BroadNet dialect set | Gulf, Egyptian and Levantine speech | 23.4 | 25.6 | 31.8 |
These are BroadNet’s lab measurements, published on its model catalogue and its Transcribe Arabic page. BroadNet ran the comparison itself, so treat the ElevenLabs and Whisper columns as one vendor’s test of its competitors, not an independent result. BroadNet also reports the model at #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard, and #1 among models whose licences permit serving through an API. On 15 November 2026 the three entries above it were Jais-ASR, Sygma-ASR-Arabic (which the board describes as beam decoding on the Cohere model) and Audar-ASR-V1-Turbo; Whisper large-v3 was #19. We have not checked each entry’s licence.
You can hear the model on your own voice before signing up: BroadNet’s browser playground transcribes up to 60 seconds of Arabic or English audio with no account. See how to use the playground.
The gap between 3.8 and 23.4 is the lesson: the same model can be near-perfect on read speech and much weaker on dialect conversation. Test on yours.
Be careful with vendor numbers in general. Many published Arabic WER figures come from the vendor’s own test, on audio you can’t see. A number is only comparable when the test set, the normalisation and the scoring tool are the same.
For the full picture, our open speech-to-text API price index lists batch and streaming prices, billing increments and Arabic support for more than 20 products, each checked against the provider’s own pricing page (CSV and JSON).
How do you check dialect coverage?
Group your test files by dialect: Gulf (Khaleeji), Egyptian, Levantine, Iraqi, Maghrebi, plus Modern Standard Arabic. Score each group on its own. If most of your customers are in Riyadh, Gulf WER matters far more than an average.
Watch for code-switching. Gulf business speech mixes in English product names and numbers. Check how numbers, dates and names come out, because those are the words your downstream system acts on. BroadNet offers a fine-tuned Arabic model for numbers, dates and names in noisy audio, on request.
What does Arabic transcription cost per audio hour?
Compare per hour of audio, because that is how your volume is measured:
| Model | Per minute | Per hour | 1,000 hours |
|---|---|---|---|
| BroadNet Transcribe Arabic | $0.0042 | $0.25 | $252.00 |
| BroadNet Transcribe (14 languages) | $0.002 | $0.12 | — |
| Soniox async (file), 60+ languages | about USD 0.0017 (derived) | about USD 0.10 | about USD 100 |
| Soniox real-time (streaming) | about USD 0.002 (derived) | about USD 0.12 | about USD 120 |
Soniox rows are from Soniox pricing, checked 14 November 2026. Soniox bills per token (about 30,000 audio tokens per hour) and quotes the hourly figures as approximations.
Soniox is cheaper per hour than BroadNet. Its file rate is less than half of BroadNet Transcribe Arabic’s $0.25 and below BroadNet Transcribe’s $0.12. Its real-time rate matches BroadNet Transcribe and still undercuts the Arabic model, and BroadNet does not document real-time streaming at all. Soniox lists Arabic (ar) among its supported languages and shows one rate for all languages. It does not publish Arabic-specific pricing, Arabic WER figures or dialect coverage, so the price is clear but the accuracy is not. Run the same dialect test on both before you choose on price.
For a support inbox, 10,000 WhatsApp voice notes of 30 seconds each cost $21.00 on the Arabic model. The math is 10,000 × 0.5 minutes × $0.0042. Run your own volume through the speech-to-text cost calculator.
When comparing providers, check the billing unit as well as the rate: per second, per minute or per started block, and whether there is a minimum charge per file. Short voice notes are where billing rounding shows up. For other providers’ per-minute rates, checked against their pricing pages, see speech-to-text API pricing compared; if you are moving off OpenAI’s endpoint, see the Whisper API alternative checklist. The Transcribe vs Transcribe Arabic comparison shows when the cheaper 14-language model is enough.
Which features should an Arabic speech-to-text API have?
Not every product needs every feature. Decide which ones matter before you compare:
| Feature | Why it matters | What to ask |
|---|---|---|
| Batch file upload | Voice notes, recordings, archives | Formats and maximum file size |
| Real-time streaming | Live captions, voice agents | Latency, and whether it’s a separate product |
| Speaker labels (diarization) | Calls with agent and customer | Batch only, or live too? |
| Word timestamps | Subtitles, search, review tools | Word or segment level? |
| Code-switching | Arabic with English terms | Does it need one language per request? |
| Data retention | Calls carry personal data | Is audio or text kept, and for how long? |
What BroadNet documents today: batch transcription through the OpenAI-compatible /v1/audio/transcriptions endpoint, WAV or MP3 input, a language field, and zero data retention. For anything else on this list, check the API reference before you rely on it.
What file limits and formats apply?
- File size. BroadNet accepts WAV or MP3 up to 26,214,400 bytes (25 MiB) per file. A larger file returns 413; split longer recordings.
- Formats. Convert .ogg voice notes and .m4a recordings first; the WhatsApp voice note guide has the ffmpeg line.
- The language field. Send
language=ar. Without it, Arabic can come back transliterated into Latin letters. - Rate limits. A pay-as-you-go key allows 30 requests a minute and 4 at once. For a large backlog, queue files and retry 429s with backoff; see errors and retries.
Is the audio stored?
Voice notes and calls carry names, phone numbers and health or banking details. Ask every provider whether audio and transcripts are retained, and for how long. BroadNet Inference does not retain either; see zero data retention and our ZDR checklist.
How do you test an Arabic transcription API in one afternoon?
- Collect 30 to 50 real clips per dialect from your production channel. Remove anything you may not share.
- Write the correct transcript for 10 clips per dialect by hand, using one agreed spelling convention.
- Transcribe every clip with each candidate API. The step-by-step Arabic guide has curl and Python code for BroadNet.
- Normalise both texts the same way and compute WER and CER per dialect. Arabic WER explained has the Python normalisation function and shows why skipping it can turn a perfect transcript into 75% WER.
- Read 20 transcripts yourself. A low WER can still hide wrong numbers or names.
How to pick clips, write references and score each dialect group is covered in detail in Arabic dialect speech recognition.
Common mistakes
- Testing only clean audio. Use real phone and voice-note recordings.
- One average WER. A good average can hide a dialect your customers use.
- Skipping the language field. Transliterated output scores as almost all errors.
- Ignoring numbers. Check digits, dates and amounts by hand; they are what your system acts on.
Where BroadNet Inference fits
BroadNet Transcribe Arabic suits batch transcription of Arabic voice notes, calls and recordings at a per-minute price, with no audio retained. The $5 starting credit covers about 1,190 minutes of Arabic audio, enough for the test above without a card. If you need features it doesn’t document, such as live streaming captions, test a provider that offers them. For Arabic text tasks after transcription, see choosing an LLM for Arabic.
