A business phone system designed around voice agents.
Voxworks builds the whole voice agent pipeline in one integrated platform: orchestration, LLM, synthesis, telephony and call flow management.
Wire it together yourself.
One voice engine. One platform. Every stage is built in.
What is a Voice Engine?
In simple terms, a voice engine simulates a live human conversation. To achieve this on a phone call without the caller noticing is no easy feat.
We must convert audio signals from the phone line into text that can be interpreted by an intelligent AI model, generate a text response and then convert that back into speech audio to be sent back over the phone line.
This entire signal processing chain must complete within a single second to sound realistic, and around half of that time is eaten up by the mobile network.
I/O channels
Speech and intelligence
Integrations
Voice gateway
Manages live sessions, routes each channel, and normalises incoming audio before it moves through the intelligence pipeline.
AI voice agents that push the envelope.
Our low-latency voice agent harness using fast inference models and industrial control techniques to maintain AI safety at speed. Architecting the full stack in this way enabled a suite of unique capabilities.
Talk to any API, mid-call.
Send a payload — including custom variables collected during the call — to any endpoint, then use the response data in later steps. Your voice agent can interact with any external API while the caller is still on the line.
# payload · live call variables { "lead": {{ custom.lead.id }}, "postcode": "3000", "cover": "comprehensive" }
Everything around the call, built into one platform.
Automations, telephony, scheduling and integrations are all native to Voxworks and managed from the same dashboard as the voice engine itself.
Automations
A lightweight n8n or Zapier built in. Pre- and post-call API calls, triggers, webhooks and email/SMS flows.
Batch Scheduler
An advanced outbound power dialer for AI agents. Long-term scheduling with quota controls, DNCR washing and self-refreshing contact lists.
Telephony
A custom-built phone switch, enabling wholesale SIP connections and phone number provisioning, native within the platform.
AI Receptionist
Set up an AI receptionist in 5 minutes. The core voice engine, tools, post-call automations and settings brought into one voice agent.
Open API
Trigger calls, upload contacts, create lists and pull post-call data.
Integrations
Book meetings into your calendar, sync leads with your CRM and process unstructured data into your systems of record, all live on a call.
Ship a calling agent in three lines.
Our open API lets you trigger calls, upload contacts, create lists and pull post-call data. The same engine that powers the dashboard, behind a clean REST interface.
await voxworks.calls.create({ agent_id: "agt_reactivation", to: "+61412000220", variables: { lead: { name: "Jordan Mei" } } });
Related reading
We publish our technical findings evaluating real-time voice AI infrastructure.
What is AI voice latency?
Why milliseconds matter in phone calls and how latency shapes caller trust.
What is voice activity detection?
How VAD decides when callers have stopped speaking and why it affects interruption handling.
Why AI voice is hard
A practical breakdown of the real-time speech stack behind natural AI conversations.
Developer FAQs
Technical Questions
On Australian mobile networks, we deliver sub-1 second "Voice-to-Voice" latency. This includes network transit, transcription, LLM token generation, and TTS synthesis. The theoretical limit is lower still with the fastest available routing, but in our architecture we utilise that extra time for safety processing and output checks.
No. Voxworks is a calibrated system and to maintain the best experience for our users we preselect the best available AI models that work for our overall stack.
We use a combination of various open source models running on fast inference GPUs to guarantee consistent outputs. In our experience, the harness does most of the heavy lifting; the intelligence of the models used in the voice pipeline should not be a differentiating factor in the consistency and quality of output for a well-defined agent.
We use an aggressive VAD (Voice Activity Detection) model on the incoming audio stream. If speech is detected while the AI is speaking, we send a "Clear Buffer" signal to cut audio instantly.
Yes. You can define "Tools" in the creation of a call script or agent and the Voice Engine will run the tool calls in parallel to the voice channel. We also use composite flows, such as the calendar booking flow, that include both voice steps and multiple tool calls that can be called as modular component within a call.
The voice engine natively supports any language however please contact support if you want to use a language other than english.
Experience the voice engine handling your real world calls.
Book in a demo and we'll wire a working agent to your use case, or jump straight into the API and build it yourself.