Arabic-first voice AI
Native-quality Arabic text-to-speech, tuned by native speakers, deployed across four regions.
Hear the range
Tap a voice. We stream a six-second sample from the same endpoint your app will call. No sign-up.
Layla
Modern Standard · Narrative
Hi, I'm Layla. A Modern Standard Arabic voice, calm and assured, built for narration, audiobooks and e-learning.How we stack up
We ship voice infrastructure for the Arabic internet, so the comparison that matters runs on Arabic content, not English-tuned demos. Numbers below are calibrated against the latest production tier each vendor offers, measured the week of launch.
| Hakim versus the leading TTS, STT, and voice-cloning APIs on Arabic content. | AzureCognitive Services | HakimBuilt for Arabic | ElevenLabsMultilingual v2 + Flash | GoogleCloud Text-to-Speech | OpenAItts-1 + Whisper | Fish Audiofish-speech v1.5 |
|---|---|---|---|---|---|---|
| Arabic excellence | ||||||
Arabic dialects with named voices Counted from each vendor's TTS catalogue, not generic 'Arabic' fallbacks. | 3 | 15 | 1 | 2 | 1 | 1 |
Arabic STT word error rate Lower is better. 7-hour Arabic-LibriSpeech eval, MSA + four dialects. | Trails | Leads | — | Trails | Competitive | — |
Voice cloning quality on Arabic How the vendor handles a 30-second Arabic sample, accent, register, dialect. | Enterprise | Native | English-tuned | Limited | — | No consent gate |
| Performance | ||||||
TTS time-to-first-byte on Arabic p50 first-audio-chunk from prompt-submit. EU-Frankfurt, 1 Gbps link, January 2026. | 350 ms | 113 ms | 320 ms | 380 ms | 700 ms | 280 ms |
Streaming TTS | WebSocket | Chunked | Chunked | Chunked | Chunked | Partial |
STT billing granularity Per-second billing means short clips don't round up to a full minute. | Per-second | Per-second | — | Per 15 seconds | Per minute | — |
| Platform & trust | ||||||
OpenAI-compatible REST API Drop-in replacement for /v1/audio/speech and /v1/audio/transcriptions. | No | Yes | No | No | Native | Partial |
Data residency regions Where your audio is processed and stored, not just billed. | EU only | GCC + EU | US only | EU only | US only | US / CN |
Voice-clone consent gate An on-platform consent step before a custom voice is callable in production. | Enterprise | Yes | Yes | No | — | No |
Hakim leads on 6 of 9 dimensions and ties on 3 more.
Methodology + sourcesNumbers from each vendor's public documentation and our 2026-Q1 internal benchmarks against their highest-quality production tier (ElevenLabs Flash, Azure Neural HD, Google Studio, OpenAI tts-1, Fish Audio real-time). Arabic STT WER measured on a 7-hour Arabic-LibriSpeech evaluation set covering MSA plus four major dialects. Vendor brand names are the trademarks of their respective owners.
Try it yourself
Direct call to the same production pipeline your app will use. 5 free generations per day, no sign-up, no card.
01 / 05
Most of the world communicates in speech. AI, until now, has assumed everyone types.
The web has grown to five billion people. Typing, in your first language, without a keyboard that fights you, has not. The gap between 'has a phone' and 'can type comfortably' is where voice belongs.
Sources: ITU global internet estimate, Ethnologue first-language speakers, regional typing-literacy surveys. Full references in the launch thesis post.
02 / 05
It's Modern Standard Arabic plus fourteen mutually-semi-intelligible regional dialects, sounds that no English-first model was trained on.
Then there's the diacritic system most real-world text omits, and a script that writes right-to-left and connects letters inside words. 'Arabic TTS' built on an English-first stack mispronounces common names by the second sentence. We built around that, not past it.
Supported range refers to current coverage in Hakim Saree' v1.3 TTS and Hakim Arab v2 STT; see the models page for the full capability matrix.
03 / 05
Arabic-first TTS engineered for streaming speed. Hakim Saree' v1.3, tuned by native speakers across 15 dialects. The first byte of streaming audio lands in roughly 113 ms p95.
Hakim Saree' (سريع, *swift*) is our streaming-speed tier · the endpoint tuned for the lowest first-byte latency. It reads Arabic script and diacritics natively, handles English inside an Arabic sentence without switching voices, and exposes the same endpoint whether you pick a region in the EU or the GCC. One API, one authentication primitive, the same billing model everywhere.
Every voice is recorded and QA'd with a native speaker before it ships. Evaluation methodology and the full three-tier capability matrix live on the models page.
04 / 05
Arabic-first streaming STT. Dialect-aware, code-switch-aware, interactive latency, first token in roughly 90 ms p95.
Hakim Arab is the half of the stack that turns voice into text. It is trained for the sentences Arabic speakers actually say: 15 Arabic dialects, English phrases inside an Arabic sentence treated as first-class input, and proper nouns + emails + numbers preserved through the transcript. 'Send the موعد to ahmed@hakim.ai' is exactly the kind of sentence that breaks English-first STT; Hakim Arab reads it straight through.
A Khaleeji voice Hakim Arab v2 is designed to transcribe. Play, then imagine the same accuracy going the other way.
No accuracy or word-error-rate claim is published on this page, we hold that conversation to the models page, where the evaluation protocol is documented.
05 / 05
Four live regions. GCC (UAE, KSA, Qatar) plus EU Frankfurt. You pick where your audio is processed; it doesn't leave that region.
Your data stays where your customers and regulators expect it to stay. Pick a region when you create a project, and every request, synthesis, transcription, storage, is handled inside that region. No silent cross-border routing, no surprises in your audit trail.
Compliant, not certified: every badge links to the exact control mapping on our security page. SOC 2 Type II audit is in progress; the Type I report is available on request.
What you can build
Text-to-speech, speech-to-text, and voice cloning, all native-quality across Arabic dialects and 30+ languages, all behind the same authentication and billing primitives.
TTS · Hakim Fast v1.3
Arabic-first streaming TTS at 113 ms p95 time-to-first-audio. MSA plus 14 regional Arabic dialects and 30+ other languages out of the box, served from your choice of four live regions.
See TTS capabilitiesSTT · Hakim Arab v2
Arabic-first streaming STT. Fifteen Arabic dialects, English and Arabic in the same sentence, 90 ms p95 first token, delivered from your choice of four live regions.
See STT capabilitiesVoice cloning
Record ten seconds, get a production voice that follows your clone's dialect and register. Cloning is instant, and custom voices share the same inference stack as Hakim Fast v1.3, same latency, same quality bar.
See voice cloningShip in minutes
Drop-in SDK for Node and Python, plus pure cURL and browser-JS snippets. Every example hits the same pipeline the paid API uses, no separate staging path.
import { Hakim } from '@hakim/sdk-node';
const client = new Hakim({ apiKey: process.env.HAKIM_API_KEY });
const audio = await client.audio.speech.create({
model: 'hakim-fast-v1',
voice: 'reem-khaleeji',
input: 'أهلاً وسهلاً بكم في حكيم.',
format: 'mp3',
});
await audio.writeToFile('hello.mp3');Built for enterprise
We operate to SOC 2 Type II, GDPR Article 28 processor, EU AI Act Article 50, UAE PDPL, and KSA PDPL requirements. The SOC 2 Type II audit is in progress; the Type I report is available on request. Every badge on this page links to the exact control mapping on the security page.
Our compliance posture today. No badge on this page implies a completed auditor attestation unless the security page's control mapping says so explicitly.
Read the security pagePricing
Free while you prototype. Pay only for what you use once you ship. Enterprise terms on request, with data residency, custom retention, and signed BAAs where they matter.
Prototype
Build against the same production pipeline the paid tiers use. Zero card, zero commitment.
No card required. Upgrade any time from the dashboard.
Ship it
The anchor plan for individual developers and small teams going to production.
Custom
Usage-based pricing, signed MSA, data-processing addenda, and the residency posture your legal team asked for.
One more thing
Five thousand free characters today. Twelve voices and every dialect we support. Thirty seconds to your first synthesis.
Start for freeNo card. Upgrade when, and if, you grow out of the free tier.