Skip to content
Ittisal
Voice AI

وكلاء صوتيون بالعربية

Arabic voice agents that answer under 800 milliseconds

Deepgram streaming recognition, GPT-4o reasoning and ElevenLabs synthesis, wired so the caller hears a response before a human would have finished reading the intent. Each tier has an in-boundary counterpart.

< 800 ms turn, p95Gulf · Levantine · EgyptianBarge-in aware
Latency

Where the 800 milliseconds go

Move the sliders to see how recognition, reasoning and synthesis compose into a single perceived turn, against a legacy IVR baseline.

Latency and cost model

Response time against unit cost

Traditional CCaaS voice stack2500 ms
Ittisal streaming pipeline780 ms

Over 8 conversational turns that is 13.8s of dead air removed from every call.

Agent-served baseline
AED 872,219
per month
With Ittisal
AED 397,296
137,500 AI minutes
Annual saving
AED 5,699,077
54% of voice handling cost

Baseline 0.95 USD per agent minute · AI pipeline 0.0095 USD per minute · currency pegs applied at par.

Interactive model. Production p95 measured from end of caller speech to first synthesised audio frame.

Pipeline

Three tiers, each swappable

Recognition

Deepgram streaming transcription with dialect models and per-speaker tagging on forked legs. Self-hosted faster-whisper for tenants that cannot egress audio.

Deepgram · faster-whisper

Reasoning

GPT-4o-mini for turn-level intent, tool calls and guardrails against tenant knowledge. vLLM-served open weights for the sovereign tier.

GPT-4o-mini · vLLM

Synthesis

ElevenLabs multilingual voices with Arabic prosody control, or Azure Neural and in-boundary TTS where residency requires it.

ElevenLabs · Azure · local TTS
Dialect

Why dialect is the whole product

A caller in Sharjah does not speak Modern Standard Arabic, and a model trained on MSA news audio will mishear them. Recognition is evaluated per dialect against tenant call recordings before a flow goes live.

Code-switching is the normal case, not an edge case. A single sentence carries an Arabic verb, an English product name and a number read in either language; the pipeline is tuned to keep all three.

  • Gulf, Levantine, Egyptian and MSA recognition profiles
  • Arabic–English code-switching inside one utterance
  • Barge-in: the caller can interrupt and the turn re-plans
  • Numbers, dates and Emirates ID formats normalised before tool calls

Hear it answer in your dialect

We run a live call on your own recordings during the demo, with recognition accuracy reported per dialect.