Blog
Thesis pieces, release write-ups, and engineering notes, written by the people building Hakim.
Three Arabic-named tiers, one API. Saree' (سريع) for sub-120 ms streaming, Faseeh (فصيح) for premium quality with non-verbal tags, Mulhim (مُلهِم) for voice generation from a free-form prompt. Plus the seven-token tag system and the new /v1/audio/voices/design endpoint.
How to transcribe Arabic with dialects and English mixed in: batch files with timestamps and speakers, realtime WebSocket streaming, and the settings that decide accuracy.
A practical guide to adding natural Arabic speech to your product: the three Hakim TTS tiers, picking a voice in the right dialect, formats and sample rates, and streaming for agents.
A blind listening test with native speakers, latency and consistency checks, and the procurement questions that matter, so you pick the Arabic voice your customers would choose.
Building an Arabic phone agent or IVR: realtime transcription of 8 kHz calls, speech at telephony sample rates, dialect-matched voices, and fallbacks callers never notice.
About ten seconds of clean speech is all a clone needs. How to record it: consent, one speaker, the right dialect and style, a quiet room and a healthy level.
Six months of releases in one place: two new voice models, cleaner realtime audio, custom vocabulary and silence trimming for speech-to-text, better voice clones, reliability and security work, and a new website.
What made long realtime TTS sessions sound wrong, how we measured it on real sessions, and the fixes now live for every voice.
How to use realtime vocabulary for brands, names and short words that are easy to mishear, and the mistakes that quietly stop it working.
Voice prompting is the capability on Hakim Mulhim v3 that turns a free-form text description into a synthetic voice. This post unpacks what voice prompting is, the three-axis prompt structure that produces consistent results, how it differs from voice cloning, and where it fits in real products.
Voice prompting works best when treated as a writing problem, not a search problem. Twelve annotated example prompts (voicebot, narration, brand voice, specialised registers), five anti-patterns that fight the model, the iteration loop that gets you from 'meh' to 'this is the one' in five minutes, and language-specific notes for English + Arabic.
Both produce a re-usable Voice row, but they fit very different jobs. Voice cloning replicates a specific consented speaker; voice design conjures a synthetic timbre from a text description. This post walks through the trade-offs, the consent and audit postures, the cost shapes, and the per-use-case decision rules we see hold up in production.
A working tutorial: joining a LiveKit room, wiring HakimSTT and HakimTTS into an AgentSession, handling interruptions correctly, and tuning turn detection for Arabic conversational cadence.
A working tutorial: assembling a Pipecat pipeline with HakimSTTService and HakimTTSService, handling interruptions and backchannel utterances, and wiring a Twilio inbound-call transport.
A practical comparison of the three names that come up in every voice-agent build conversation, the volume threshold that should drive your choice, and why the STT/TTS provider's dialect coverage matters more than the framework once you're outside English.
GPT-Realtime-2 and Gemini 3.1 Flash Live make a compelling pitch for end-to-end speech-to-speech models. Here's why cascaded architecture still wins on dialect precision, auditability, and vendor flexibility for Arabic voice products in 2026.
Why 'the AI keeps talking over me' is a plumbing problem, not a model problem, why English-tuned turn detectors misread Arabic conversational cadence, and what a TTS layer needs to cancel cleanly within 100 ms.
Acknowledgment phrases, parallel function calls, timeout fallbacks, and interruption policy for side-effecting tool calls, the patterns that keep a voice agent talking while a backend API is still thinking.
The GPAI model obligations landing in August are a different deadline than the Article 50 transparency rules. What a voice-agent operator specifically needs on file, and the logging checklist regulators will actually ask for.
Containment rate is the metric every vendor quotes and few buyers know how to interrogate. How the number gets inflated, the dialect variable most quotes don't disclose, and what to measure yourself before it drives a staffing decision.
How to evaluate Arabic TTS, STT, and voice cloning APIs without the marketing claims. Eleven criteria that actually matter, what to ignore, and how the major providers stack up across MSA and the regional dialect tree.
Building Arabic-language voice agents for the Gulf, Egypt, and the Levant. STT dialect coverage, sub-120 ms TTS, interrupt handling, and the four mistakes that sink voice-agent rollouts in MENA.
AI dubbing isn't just translation plus TTS. A breakdown of the seven-stage pipeline, where dialect choice changes everything, lip-sync constraints, and how studios actually integrate AI dubbing into existing post-production workflows.
Khaleeji isn't one dialect, it's a family. A primer on Najdi, Hejazi, Emirati, Kuwaiti, Qatari, and Bahraini varieties; what voice AI actually needs to handle them; and what 'covered' looks like in production.
Most voice AI treats Arabic as an afterthought. Here's why we rebuilt the stack with MSA and fifteen dialects as the default, not the fallback.
Behind the sub-120 ms TTFB number: chunked autoregressive decoding, WebSocket transport semantics, mid-stream cancel, and the architectural choices that separate a streaming TTS from a batch TTS labelled streaming.
A deep dive on the latency budget: model surgery, custom inference, and the four-region topology that gets a synthesis call to your user in 113 milliseconds.
Fifteen Arabic dialects and six MENA language tiers, evaluated on regional corpora, not press releases. How we measure it, and how we plan to grow it.
What you're actually paying for: per-character vs. per-second, free-tier traps, voice-cloning amortisation, dialect surcharges, and the hidden costs that show up at 100M characters per month.
The EU AI Act came into force on 2 February 2025. What changed for voice cloning specifically: transparency disclosures, deepfake labelling, the GPAI obligations, and how the EU regime intersects with the GCC's identity-misuse statutes.
Frankfurt, UAE, KSA, Qatar, four full deployments from day one. Why the hub-and-spoke model doesn't work for Arabic voice AI, and what we built instead.
Eighteen months head-down, four founders still on the stack, a production TTS + STT model family, and four live regions. Today we come out of stealth, here's the arc.
Most MENA-targeted products are bilingual by default. How to design a voice layer that handles AR/EN code-switching gracefully: language detection, prosody at the boundary, voice catalogue ergonomics, and the LLM prompt patterns that don't break.
Quality is a single axis. Voice cloning is a four-axis problem: quality, consent, traceability, revocation. Here's how we built each of them, and what we refuse to do.
A retrospective on the nine weeks of decoder surgery, the Beirut-Cairo-Riyadh corpus collection, and the boring infra work that separated a notebook demo from a production API.