Voice & Video Practice

Voice & Video AI agents.

Sub-500ms bilingual voice agents, dialect-aware ASR, and photoreal video presenters — deployed into Genesys, Avaya, Twilio and branch kiosks across the Gulf.

Talk to a leadAll services
Practice
Voice & Video Practice
Delivery
Riyadh · London · Lahore
Model
Fixed-scope · Retainer
Timeline
6–16 weeks / phase
01 — What we do

Three pillars that carry the practice.

01

Real-time streaming voice

Sub-500ms end-to-end Arabic + English voice, streamed over SIP into your existing telephony — no rip-and-replace.

02

Photoreal avatars

On-brand video presenters for training, onboarding, branch kiosks and marketing — lip-synced to Arabic and English scripts.

03

Contact-centre co-pilots

Live agent-assist with real-time summaries, next-best-action, sentiment tracking and automated QA scoring on 100% of calls.

“
We ship this practice as a small, opinionated system — running infrastructure, evaluation harnesses, and the rituals that make leadership trust the numbers.
02 — Capabilities

What we build inside this practice.

Every engagement is scoped as a small system, not a slide deck. You get running infrastructure, documentation, and the metric-store hooks to measure it.

Scope an engagement
01Streaming Arabic + English voice
02Gulf / Levantine / MSA ASR
03Telephony (Genesys · Avaya · Twilio)
04Photoreal avatar generation
05Live agent-assist & QA
06Call summarisation & tagging
07IVR replacement
08Branch kiosk voice UX
Reference architecture

How the pieces move.

TelephonySIP · WebRTCASRdialect-awareAgentplanner · toolsTTSvoice · avatarQAscore · coach
03 — In production

Numbers from the field.

450ms
End-to-end voice latency

Self-hosted deployment for a KSA bank, on-prem GPU inference.

38%
AHT reduction

Live agent-assist on inbound Arabic + English queues.

24/7
Bilingual coverage

Without expanding contact-centre headcount.

04 — The problem we solve

English-first voice stacks break on Gulf Arabic.

Off-the-shelf ASR trained on MSA and English hits 60–70% accuracy on Khaleeji dialect — unusable in a bank IVR. We build the whole pipeline: dialect-tuned ASR, streaming planners, TTS that doesn't sound robotic, and telephony integration that survives call transfers, holds and codec renegotiation.

96%
Arabic ASR accuracy on Gulf-dialect calls
<500ms
P95 barge-in latency, end-to-end
100%
Calls transcribed, tagged and QA-scored
SIP
Native integration — no telephony rip-and-replace
05 — How we deliver

A methodology we've run — not a slide.

Every engagement follows the same phased spine so leadership always knows what ships in the next two weeks.

  1. 01
    Phase 1

    Call taxonomy

    Sample 500+ calls, define intents, dialects, escalation paths and success metrics.

    Intent taxonomyDialect mixSuccess KPIs
  2. 02
    Phase 2

    Model pipeline

    Tune ASR on Gulf dialect, wire planner, select TTS voice, benchmark latency.

    Tuned ASRPlanner graphTTS voice pack
  3. 03
    Phase 3

    Telephony integration

    SIP / WebRTC bridge into Genesys, Avaya or Twilio; codec, DTMF, transfer handling.

    SIP bridgeCall-control testsFailover plan
  4. 04
    Phase 4

    Rollout & QA

    Progressive rollout with live agent-assist first, then autonomous handling on low-risk intents.

    Assist rolloutQA scorecardsCoaching loop
06 — Reference stack

Vendor-neutral. Battle-tested.

ASR (speech in)
Whisper large-v3AssemblyAIDeepgram Nova-3Custom Arabic fine-tunes
TTS (speech out)
ElevenLabsCartesia SonicAzure Neural ArabicCustom voice cloning
Realtime & agents
OpenAI RealtimeLiveKit AgentsPipecatLangGraph
Telephony & video
TwilioGenesys Cloud CXAvaya OneCloudWebRTCHeyGen avatarsD-ID
07 — What you get

Concrete deliverables.

  1. 01Call taxonomy and intent library
  2. 02Gulf-dialect ASR benchmark and tuned model
  3. 03Streaming voice agent with barge-in
  4. 04TTS voice pack (AR + EN, brand-tuned)
  5. 05SIP / WebRTC telephony bridge
  6. 06Live agent-assist UI for the contact centre
  7. 07Automated QA scoring on 100% of calls
  8. 08Photoreal avatar library for branch / training
08 — Buyer questions

What leadership actually asks.

Do you support Gulf, Levantine and MSA?

Yes. Our ASR is tuned across Gulf (KSA, UAE, Kuwait, Bahrain, Qatar), Levantine and MSA, with dialect classification on the front so the planner routes to the right voice and script.

Can you deploy on-premise for regulated banks?

Yes. The full stack — ASR, planner, TTS — runs on-prem on a single-node GPU. See our Edge Inference practice for hardware sizing.

How do you handle call transfers and holds?

The SIP bridge supports blind and warm transfer, hold-music injection and codec renegotiation. We test the failure modes explicitly during telephony QA.

Is avatar video too gimmicky for enterprise?

Depends on the surface. In branch kiosks, onboarding videos and internal training, avatars cut production cost dramatically. In outbound customer video calls, we usually recommend against them today.

Latency numbers — real or lab?

P95, measured on production traffic with real SIP legs. Lab numbers are always faster and we don't quote them.

Ready to scope this into your stack?

Book a working session with a lead architect.

Book a session See our research
Book a Demo