FAQ
What is this demo?
This is a live demonstration of Voxworks AI-powered voice technology. When you click “Talk to Us,” you’ll be connected to an AI voice agent that simulates a real insurance company receptionist in real time. This is an early preview of a "natural IVR", that effectively eliminates an IVR menu system. The idea is to directly route the customer to the information, system process or human contact agent based on a natural conversation.
How does it work?
This demo uses WebRTC to establish a live 2-way audio connection directly in your browser. The audio is connected to our voice engine, which processes the recording from your microphone and converts into text using a speech-to-text AI model. From there, the conversation is modelled by a LLM assessing your responses against the designated script, which sends back a response within our predefined guardrails. The response message is then synthesized into audio using a text-to-speech model trained on an Australian accent, and sent back to the device. This all happens within ~500-700ms, fast enough so that you can't notice any latency, and all run on servers and GPUs in a data centre in Sydney.
What does the call script look like?
Within our web app, a call script be designed by simply editing what you want the AI to say at each conversational step. Steps can be connected to subsequent steps with conditions that are assessed in real time based on what the contact says. Steps can be grouped into conversational flows and then reused for other call scripts.
How is this different to other voice AI I've experienced?
Our core IP is the model orchestration layer which drives the content of the conversation and stitches everything together in real time. Most other AI Voice platforms use a single long prompt to generate every message. We find this approach doesn't provide acceptable control over model outputs and results in hallucinations. To solve this we use a different prompt for every turn of the conversation. The system traces the conversation through a predefined pathway using AI. This approach also provides granular control over system settings depending on the requirements of each step, such as modifying AI model parameters (jump-in speed, patience, thinking time, etc), providing tool access to the AI only once it has reached a predefined step, or capturing structured data within a specific step.
What compliance features are included?
Whenever a call script is modified and saved in the system, it is assessed against your predefined policy rubric by an AI agent. The team admin can control whether calls can be placed that do not comply with predefined policies. The compliance module can also be applied to prevent non-compliant model outputs during live calls.