All insights

ยท 7 min read

AI Voice Agent in a Box: A Private Voice Agent That Lives in Your Office

Voice Agent in a Box runs an AI voice agent, a local LLM and a config UI on an NVIDIA DGX Spark in your office. Arabic and English, no token rent.

DHIA

Short answer: An AI voice agent is software that listens to a caller, understands the request, and answers out loud in real time. DHIA Voice Agent in a Box packages an AI voice agent, a local large language model (LLM) and a Voice Agent Config UI inside one NVIDIA DGX Spark that sits in your building. Calls are handled by AI you own, with no per-token bill and no outside model provider in the loop.

Every customer call carries data: names, phone numbers, account details and sometimes health or financial information. When we showed Voice Agent in a Box at the AI Everything exhibition in Dubai, demand from visitors was very high. This guide explains what an AI voice agent is, why running one on premises matters, and how to evaluate one before you buy.

What is an AI voice agent?

An AI voice agent is a conversational system that handles spoken conversations end to end. Most production designs use a cascaded pipeline: automatic speech recognition (ASR) turns the caller's speech into text, a large language model decides what to say or do, and text-to-speech (TTS) speaks the reply. NVIDIA describes this pattern in its open Nemotron Voice Agent Blueprint, which integrates ASR, LLM and TTS and adds voice activity detection and end-of-utterance logic so the agent knows when to start and stop speaking, including when a caller interrupts (NVIDIA Nemotron Voice Agent Blueprint).

The same blueprint lists DGX Spark among its supported single-GPU local deployment targets, which is the practical point: a complete voice pipeline no longer needs a data center.

An AI voice agent is different from a traditional IVR menu. Instead of "press 1 for billing", the caller simply says what they need, in their own words, and the agent answers from the knowledge you have approved.

What is DHIA Voice Agent in a Box?

Voice Agent in a Box is DHIA's "AI in a box" voice offering. It has three parts, all running inside a single NVIDIA DGX Spark on your premises:

  • An AI voice agent that holds the spoken conversation.
  • A local LLM that reasons over your approved business knowledge and decides the response.
  • A Voice Agent Config UI where your team sets up and adjusts how the agent behaves.

It follows the same ownership model as the rest of DHIA. The hardware sits in your office, on your network. There is no monthly token rent, so a busy month of calls does not create a bigger AI bill. DHIA works natively in Arabic and English, and its Arabic customer service use case is built around private voice and chat agents that answer customer questions from approved knowledge and keep call transcripts inside your environment.

Calls reach the box over standard telephony. A business can put a local mobile SIM into a GSM gateway connected to its PBX, which routes calls to the AI voice agent. Or it can get a SIP trunk account from its local telecom provider and connect that instead of a GSM gateway.

Why run a voice agent on premises?

When a voice agent runs as a hosted service, caller audio and transcripts are processed on infrastructure you do not control. An on-premises voice agent changes three things.

  1. Data location. Audio, transcripts and the knowledge the agent uses stay on hardware in your building. If you are new to the idea, our guide to on-premises AI explains how on-premises, self-hosted and cloud AI differ.
  2. Cost shape. You own the capacity. You plan for it like infrastructure, not like a metered utility that grows with every minute of conversation.
  3. Control. Your team configures the agent in the Voice Agent Config UI, and DHIA's retention controls let you decide how long conversations and logs are kept.

Voice data and GCC data protection rules

Call recordings and transcripts usually contain personal data. In the UAE, Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data came into force on 2 January 2022. It prohibits processing personal data without the owner's consent except in specified cases, and it sets out requirements for cross-border transfer and sharing of personal data (UAE Government portal, Data protection laws).

In Saudi Arabia, the Regulation on Personal Data Transfer outside the Kingdom requires controllers to limit transfers to the minimum necessary, and to conduct a risk assessment for transfers made under appropriate safeguards or exemptions and for continuous or large-scale transfers of sensitive data (SDAIA, Regulation on Personal Data Transfer outside the Kingdom).

Running the voice pipeline on your own hardware means call audio does not have to leave your premises for the agent to work, which simplifies the transfer analysis. It does not remove your other obligations, such as notice, consent and retention. Treat this section as context, not legal advice, and review your setup with your compliance team.

The hardware: NVIDIA DGX Spark

Voice Agent in a Box runs on the NVIDIA DGX Spark. These figures come from NVIDIA's own product page (NVIDIA DGX Spark):

Specification NVIDIA DGX Spark
Chip NVIDIA GB10 Grace Blackwell Superchip
AI performance Up to 1 petaFLOP at FP4 precision
Memory 128 GB coherent unified system memory (a 64 GB option is sold through OEM partners)
Model size for inference Models up to 200 billion parameters on one 128 GB unit
Scaling ConnectX networking links up to four DGX Spark systems
Size and weight 150 mm x 150 mm x 50.5 mm, 1.2 kg
Power supply 240 W

NVIDIA positions DGX Spark as a platform for always-on agent workloads that run locally and reduce the need for cloud-based token generation. In plain terms, it is a desk-sized box that is powerful enough to run a voice agent and its language model without sending anything to a cloud API.

What a voice agent can take off your team's plate

The best first use cases are high-volume calls with answers that already exist in your documents:

  • Answering repeat customer questions about opening hours, locations, services, documents required and order or application status, from approved knowledge only.
  • Bilingual front desk coverage in Arabic and English, including after hours.
  • Capturing caller details and the reason for the call, then handing a clean summary to the right person.
  • Keeping transcripts inside your environment so managers can review calls and improve answers.

When the agent cannot answer safely, DHIA's design is to hand off to a manager, department owner or existing ticketing workflow instead of guessing.

How to evaluate any AI voice agent

Whether you choose DHIA or another vendor, ask these questions before you sign:

  1. Where is caller audio processed and stored, and can you point to the physical hardware?
  2. How does it handle your callers' real Arabic, including dialect and mixed Arabic and English? Test it with recordings of your own calls.
  3. What is the response delay on a live call in your environment, not in a demo video?
  4. What happens when it does not know the answer?
  5. Who can change its behavior, and how quickly? Can a manager do it, or does every change need a developer?
  6. What is the monthly cost at three times your current call volume?
  7. Can you export or delete transcripts on your own schedule?

Getting started

DHIA deployments start with one discovery conversation to find the bottleneck costing you the most, then one focused agent, then scale. For voice, a sensible first step is to pick one call type, such as customer FAQs or appointment questions, agree the approved answers, and test the agent with your own team before customers hear it.

Request a private demo to see Voice Agent in a Box with your own call scenarios, or email info@raqi.ai.

Frequently asked questions

What is an AI voice agent?

An AI voice agent is software that holds a spoken conversation in real time. It converts speech to text, uses a language model to decide the reply, and speaks the answer back with text-to-speech.

What is in DHIA Voice Agent in a Box?

An AI voice agent, a local LLM and a Voice Agent Config UI, all running inside one NVIDIA DGX Spark installed on your premises.

Does Voice Agent in a Box send call audio to the cloud?

No outside model provider is needed to run it. The voice agent and the language model both run on the DGX Spark in your building, so audio and transcripts can stay inside your environment.

Does it speak Arabic?

DHIA works in Arabic and English. Test any voice agent with recordings of your own callers before you commit.

Is there a per-minute or per-token fee?

DHIA has no token rent. You own the hardware and the capacity. For a quote, request a demo or email info@raqi.ai.

Sources

Request a demo