Best Arabic speech-to-text for Saudi dialects: how to choose
An honest comparison framework for Arabic STT in Saudi Arabia — dialect and code-switching accuracy, telephony audio, custom vocabulary for Saudi names, streaming latency, and how to benchmark on your own calls.
Speech-to-text is where an Arabic voice deployment is won or lost. Everything downstream — the model's answer, the CRM write, the transfer decision — inherits whatever the transcriber heard. And Arabic is genuinely harder than English here: dialect diverges sharply from the written standard, speakers code-switch constantly, and telephony audio strips exactly the frequencies that distinguish similar consonants.
The landscape, honestly
The market splits into general-purpose multilingual engines, telephony-tuned engines, and Arabic-focused research models. Published benchmark numbers are close to useless for this decision, because almost all of them are measured on clean read speech rather than dialectal conversation over an 8 kHz phone line. Treat the table below as a map of what to test, not a ranking.
| Category | Examples | Strength | Watch for |
|---|---|---|---|
| Large multilingual engines | OpenAI Whisper and its derivatives, Google, Microsoft Azure, AWS | Broad language coverage, strong on Modern Standard Arabic, easy to trial. | Dialect accuracy drops off; check streaming support and where audio is processed. |
| Telephony-tuned engines | Deepgram, AssemblyAI and similar real-time providers | Built for narrowband audio and low-latency streaming, good diarization. | Arabic is rarely their flagship language — verify dialect handling specifically. |
| Arabic-focused models | Regional and academic Arabic ASR efforts, dialect-specific fine-tunes | Much better on dialectal speech when the dialect matches the training data. | Operational maturity, uptime, and support vary; often needs self-hosting. |
| Self-hosted open weights | Whisper-family and open Arabic fine-tunes on your own infrastructure | Full control over residency and cost at volume; fine-tunable on your calls. | You own scaling, latency engineering, and ongoing accuracy work. |
What to measure
Dialect word error rate, not headline WER
A vendor's Arabic WER is almost always measured on MSA. Build your own test set from real Najdi, Hijazi and Khaleeji calls and measure separately per dialect. The spread between MSA and dialect performance on the same engine is frequently larger than the spread between engines.
Entity accuracy, weighted heavily
Overall WER hides the errors that actually cost you money. Score these separately: Saudi mobile numbers, SAR amounts, national ID and iqama digits, compound Arabic names, city and district names, and your own product names. A single wrong digit in a phone number fails the call no matter how good the rest of the transcript is.
Code-switching
Test utterances that switch language mid-sentence. Some engines require a language hint per stream and will mangle the switch; others detect it. This is not an edge case in Saudi business calls — it is the norm.
Telephony conditions
Benchmark on 8 kHz audio with real background noise, not on clean recordings. If your evaluation set was recorded on a laptop microphone in a quiet room, your results will not predict production at all.
Streaming latency and endpointing
For live calls you need partial results fast and, just as importantly, correct endpointing — knowing when the caller has finished speaking. Aggressive endpointing interrupts people mid-thought; lax endpointing makes the agent feel slow. Measure both time-to-first-partial and how often the agent cuts a caller off.
Levers that beat switching vendors
- Custom vocabulary or keyword boosting for your product names, branches, and common Saudi given names — usually the single largest accuracy win available.
- Constrained parsing for digits: capture phone numbers and IDs against a known format rather than trusting free-form transcription.
- Confirmation turns for high-cost entities — read the number back once, and only once.
- Fine-tuning on a few hours of your own labelled calls, if your volume justifies it.
- Feeding conversation context to the transcriber where supported, so it expects the vocabulary of the current step.
A benchmark you can trust
Take one hundred real calls, spread across dialects and across your busiest intents. Have a bilingual reviewer produce reference transcripts. Run every candidate engine over identical audio. Report three numbers per engine: overall WER, entity accuracy on the list above, and median time-to-first-partial. Anything less than this and you are choosing on marketing.
Voho runs this benchmark for the deployments we operate and re-runs it whenever a provider ships a model update, because engine rankings do change. If you would rather not build the harness yourself, that is a reasonable thing to hand to an operator.
Frequently asked
- Which speech-to-text engine is most accurate for Saudi Arabic?
- It depends on your audio and vocabulary, and it changes as vendors ship updates. Engines that lead on Modern Standard Arabic often do not lead on Najdi or Hijazi conversation over a phone line, so the only reliable answer comes from benchmarking candidates on a hundred of your own recorded calls with reference transcripts.
- Does Whisper work well for Saudi dialects?
- Whisper-family models are a strong general starting point for Arabic and are easy to self-host for data-residency reasons, but accuracy falls on dialectal conversational speech compared with Modern Standard Arabic. Fine-tuning on your own dialect data and adding keyword boosting for Saudi names and product terms typically matters more than the base model choice.
- How do I handle Arabic-English code-switching in transcription?
- Choose an engine that detects language changes within an utterance rather than requiring a fixed language per stream, and include mid-sentence switches in your test set. Adding your English product and job-title vocabulary to a custom vocabulary list also reduces the errors that cluster around switch points.
- Why is my transcription accuracy worse in production than in testing?
- Almost always because testing used clean wideband audio and production runs on narrowband telephony with background noise, or because the test set was Modern Standard Arabic and real callers speak dialect. Rebuild the evaluation set from actual recorded calls and the gap usually explains itself.
Keep reading