Research at McMaster University
A voice companion for older adults, connected to their care team.
Everyday help by voice or chat, with care teams able to follow along when it matters. Built to feel human, stay clear, and protect people's privacy.
~250 ms
average voice response
Self-hosted
open-source stack
2 papers
peer-reviewed so far
Multimodal chat
Hybrid voice and text conversations
Older adults interact through voice or text in the Care Companion app. An ultra low-latency agentic AI stack (about 250 ms average voice response) handles each turn with natural overlap: adaptive interruption handling, an on-device end-of-turn model so the agent does not cut in too early, and adaptive voice speed for clarity.
How it works
From real-time agent to clinical workflow
How the voice-and-video care companion connects patient experience, intelligent workflows, and secure provider tooling in one line of sight.
Privacy & ethics
Designed for consent, oversight and accessibility
HIPAA-aligned consent and oversight, traceable AI outputs, and encrypted sessions with no unrelated third-party analytics or ads. Live transcription supports people when hearing is limited.
Privacy & consent
Dual patient consent, encrypted transit and at-rest data, audit logging, and no sale of conversation data to third parties.
Open & self-hostable stack
Components and models are composed from open-source, self-hostable software so PHI and workloads can stay on infrastructure you control—not opaque SaaS you cannot audit.
Transparency for clinicians
Dashboard summaries link back to source messages so providers can verify AI-generated text.
Human in the loop
The chatbot does not diagnose; escalations and provider review keep decisions with care teams.
Voice & trust
Built for natural, private conversations
Voice and avatar tuned for clarity, with guardrails and encryption so sessions stay in environments you control, on open, self-hostable services suited to regulated care.
Adaptive interruption handling
Smooth barge-in and overlap so users are not locked into rigid turn-taking when they think aloud or self-correct.
Ultra realistic avatar interface
A lifelike avatar carries tone and attention—often easier to follow than text-only or flat UI for extended sessions.
Ultra low-latency voice agent
About 250 ms average response time so replies land in the same breath as natural conversation.
Intelligent end-of-turn
On-device end-of-turn detection cuts down on clipped sentences and premature agent talk-overs.
Adaptive voice speed
Playback pacing adjusts for clarity—slower when needed, without sounding robotic or rushed.
Secure & private connection
Encrypted sessions with no third-party data sharing for unrelated analytics, ads, or model training—aligned with stacks you can self-host when policy requires it.
Publications
Research behind the project
Peer-reviewed work on generative AI in senior care, from system design to ethical judgment.
Help shape care technology for older adults
We're looking for older adults, caregivers and clinical partners to try the companion and tell us what matters.