Arabic Dialects · Speech to Text

Saudi · سعودي

Saudi Arabic
Speech to Text

A speech-to-text model trained specifically on Saudi Arabic, covering all four regional varieties and the Arabic-English code-switching that defines professional Saudi speech in the Vision 2030 era. Real accuracy. Real dialect.

dialect: "saudi"30M+ native speakers · real-time · code-switching

30M+

Native Saudi Arabic speakers

<18.6%

WER on Saudi colloquial speech

<300ms

Streaming latency

4

Saudi sub-dialects supported

Try it

Try Saudi Arabic Recognition

Record yourself speaking Saudi Arabic and watch the words appear as you talk. The demo runs Hamsa's full production STT, the same model your API calls will use.

Speech to Text · speech in → text out

No sign-up
Microphone
press to record

Your words will appear here.

Press record and speak Arabic or English. The words appear as you talk.

Free to start · in your browserTranscribe in Hamsa Media

Saudi Arabic Overview

What is Saudi Arabic?

Saudi Arabic is spoken by 30M+ native speakers across four regional varieties: Najdi, Hejazi, Southern, and Northern. It is the dominant media dialect of the Gulf. Professional contexts feature ~25% Arabic-English code-switching, a defining characteristic of modern Saudi business speech that generic models consistently fail to handle.

Key features that distinguish Saudi Arabic from Modern Standard Arabic:

  • ج (jeem) pronounced as 'y' across wide regions
  • ق (qaf) varies between preservation and glottal stop by region
  • ~25% Arabic-English code-switching in professional/tech speech
  • Specialized vocabulary tied to the Vision 2030 economy

Native speakers

30M+

Saudi Arabia

Sub-dialects

4

Najdi · Hejazi · Southern · Northern

Code-switching rate

~25%

In professional/tech speech

WER

<18.6%

Hamsa benchmark 2024

The problem

Why generic Arabic STT fails
on Saudi Arabic

A model trained on MSA reaches 42–65% word error rate on real Saudi conversational audio. Here is why.

MSA models fail on Saudi colloquial

Models trained on MSA register 42–65% WER on real Saudi conversational speech, unusable for any consumer-facing product.

Mixed-language speech breaks generic models

Saudi professionals mix Arabic and English naturally in almost every business conversation. Models that can't handle this lose significant meaning in technical exchanges.

Regional diversity compounds errors

The phonological distance between Najdi and Hejazi alone is enough to confuse non-specialized models.

Transcription lag kills live applications

In real-time environments, every second of delay means worse UX and higher operational cost.

The solution

How Hamsa handles
Saudi Arabic natively

✓

Built for Saudi colloquial from the ground up

Hamsa trains on hundreds of thousands of hours of real Saudi speech, not an MSA adaptation, but a purpose-built model for Saudi Arabic in all its varieties.

✓

Advanced code-switching recognition

Handles mid-sentence Arabic-English switching, producing correctly scripted output for each language.

✓

All four sub-dialects in one model

No multi-config setup: Najdi, Hejazi, Southern, and Northern covered by a single API call.

✓

Real-time streaming under 300ms

Optimized infrastructure for contact centers, voice assistants, and live interactive applications.

Capabilities

Everything you need for
Saudi Arabic STT

✓

Four sub-dialect coverage

Najdi, Hejazi, Southern, Northern, one model handles the full range of Saudi regional speech.

✓

Code-switching support

Recognizes natural Arabic-English mixed sentences as Saudis use them in real work environments.

✓

Real-time streaming

Sub-300ms latency for contact centers and voice assistants.

✓

Speaker diarization

Identifies and labels multiple speakers in calls and meetings.

✓

Automatic punctuation

Produces punctuated, production-ready transcripts without manual editing.

✓

Noise robustness

Trained on contact center and public environments, handles background noise without accuracy loss.

Transcript examples

Real Saudi Arabic.
Accurate transcription.

Contact center (Najdi)

هلا والله، أبشر طال عمرك، بشيّك على الطلب حقك الحين وأرد عليك.

"Hello there. Certainly, sir, I'll check your order right now and get back to you."

Najdi: "أبشر" (certainly), "طال عمرك" (polite address), "حقك" (your), "الحين" (now)

Customer service (Hejazi)

إيش اللي صار؟ أنا كلمتكم أمس وقالولي خلاص انحلت، بس لسه المشكلة زي ما هي.

"What happened? I called you yesterday and they told me it was fixed, but the problem is still the same."

Hejazi: "إيش" (what), "لسه" (still), "زي ما هي" (just as it was)

Business meeting

خلونا نسوي meeting بكرة ونراجع الـ deadline حق المشروع، وش رايكم؟

"Let's hold a meeting tomorrow and review the project deadline. What do you all think?"

Code-switching: "meeting" and "deadline" mid-sentence, "نسوي" (do), "وش" (what)

Use cases

Where Saudi Arabic STT
delivers results

Vision 2030 & Government

Absher and Nafath voice service transcription

Transcribing e-government platform calls with accurate recognition of regional Saudi speech patterns across all citizen queries.

Contact Centers & Retail

Saudi contact center automation

Processing thousands of calls daily, recognizing regional accent variation among agents and customers across the Kingdom.

Fintech

Financial services compliance transcription

Documenting advisory calls for regulatory compliance, with support for Islamic finance terminology and investment product vocabulary.

Media & Content

Efficient local content production

Auto-transcribing Saudi podcasts, broadcasts, and digital content for captions and searchable text outputs.

Accuracy benchmark

Saudi Arabic Speech Recognition Benchmark

Tested on hours of real Saudi colloquial speech from contact centers, informal conversation, and podcasts. Under 18.6% WER, compared to 42–65% for generic MSA-trained models on the same input.

Word Error Rate (WER) · lower is better

Hamsa STT (Saudi)us18.6%
Generic Arabic STT42%
MSA-trained model55.3%
GPT-4o-transcribe58.1%

Measured on Saudi colloquial Arabic benchmarks · 2024 · Lower is better

Architecture

How it works

01

Audio input

Send audio to the Hamsa API, streamed live or uploaded as a file. All common formats accepted.

02

Dialect detection and routing

The model identifies Saudi Arabic phonological patterns and code-switching, routing to the appropriate recognition engine.

03

Text output

Returns clean Arabic text with punctuation, timestamps, and speaker diarization metadata.

Developer API

Two lines to Saudi Arabic speech recognition

Drop in the Hamsa SDK, set dialect: "saudi", and start streaming transcripts. No preprocessing pipeline, no normalization to MSA, no per-region configuration.

Read the API docs →
saudi-stt.tsstreaming
import { Hamsa } from '@hamsa-ai/sdk';

const stream = hamsa.stt.stream({
  model:       'hamsa-stt-s4',
  dialect:     'saudi',
  diarization: true,
});

stream.on('transcript', (t) => console.log(t.text));
// → "وش الأخبار؟ كيفك اليوم؟" · 00:01.240

Related services

More Hamsa for Saudi Arabic

Why Hamsa

Built for Arabic from the ground up

✓

Native, not adapted

Hamsa trains on real Saudi speech, not English or MSA models adapted for Arabic. That is how it stays under 18.6% WER where MSA-trained models reach 42–65%.

✓

Enterprise-grade SLA

99.9% uptime, SOC 2 Type II compliance, and data residency in the MENA region. Your audio stays in-region.

✓

Fastest Arabic streaming

Under 300ms streaming latency in production. Real-time transcription that keeps pace with human speech rates.

✓

13+ dialects, one API

Switch dialect with a single parameter. Run A/B tests across Saudi and other Gulf dialects. No separate integrations, no separate contracts.

FAQ

Saudi Arabic STT FAQ

Does it support different Saudi regional dialects?

Yes. Trained on Najdi, Hejazi, Southern, and Northern speech, no per-dialect configuration required.

How does it handle Arabic-English code-switching?

Handles natural code-switched speech as Saudis use it, outputting correctly scripted text for each language.

How accurate vs. generic models?

Under 18.6% WER on real Saudi speech, versus 42–65% for MSA-trained models facing Saudi colloquial input.

Suitable for real-time contact centers?

Yes. Sub-300ms streaming latency, designed for contact center and voice assistant environments.

Can I add custom vocabulary?

Yes. Supply a custom word list with product names, technical terms, and sector-specific vocabulary.

Explore more

13+ Arabic dialects, supported

Partner ecosystem

Trusted Across the
Technology Ecosystem

Hamsa works with leading technology companies, integrators, and solution providers to bring enterprise Voice AI to organizations across the region.