The most humanistic AI for high-stakes conversations.
TalkLab builds voice AI role-play partners with a working model of emotion. Organisations use them to train people, simulate the conversations that go wrong, and assess performance at scale.
- Real-time voice
- Emotion and rapport engines
- Founded at the University of Toronto
Animated example: an AI persona playing a worried family member at a hospital. Its emotional state moves from angry to relaxed as a trainee acknowledges them and gives clear information.
Pilots and traction
- Conversations
- 10K+
- Hours of conversation
- 500+
Ongoing pilots
Demos heading to pilot
Voice simulations for the conversations that can't go wrong.
Role-play used to need an actor, a trainer and a free afternoon. An AI persona plays the other person on demand, stays in character, reacts to how it's spoken to, and debriefs at the end.
- Frontline healthcare
Anxious patient
- Trains
- Difficult news, informed consent, bedside manner
- Measures
- Empathy, clarity, teach-back
- Public safety
Distressed caller
- Trains
- De-escalation and crisis communication
- Measures
- Pace, rapport recovery, offering choices
- Corporate L&D
Hiring lead
- Trains
- Interviews, difficult feedback, client calls
- Measures
- Talk-time share, follow-up questions, objection handling
- Higher education
Oral examiner
- Trains
- Oral exams and viva practice
- Measures
- Rubric grades straight to the gradebook
Teach the skill, practise it, then measure it.
One platform replaces the expert who can only be in one room at a time. Scroll through a de-escalation simulation built for public-safety training.
- 01 · Teach
An expert instructor that never runs out of time.
An AI instructor teaches the framework first: domain knowledge, examples and the moves that experts actually make.
- 02 · Practice
A conversation partner that reacts like a person.
Each persona has its own personality, goals and emotional state. Say the wrong thing and it pushes back. Build rapport and it settles.
- Anxious patient
- Hiring lead
- Judging panelist
- Stakeholder
- 03 · Measure
Evidence of skill, with feedback people can use.
The expert assessor grades each conversation against your rubric and writes feedback. Grades roll into a gradebook ready for your LMS.
Every persona carries an emotional state.
The engine places each persona on Russell's circumplex of affect. What the trainee says moves it, and the persona's replies follow. Drag the point to feel the difference.
Neutral · valence +0.0 · arousal +0.0
- Valence
- Pleasant states read as rounded, unpleasant ones as pinched. People match round forms to soft sounds and sharp ones to tense sounds.
- Arousal
- Sets the breathing, from 6 breaths a minute when calm to 20 when activated.
- Intensity
- The further from neutral, the more the contour moves.
Russell (1980); Köhler (1929); Ramachandran & Hubbard (2001); Boiten, Frijda & Wientjes (1994).
Built for training teams that run at scale.
Real-time voice
Natural turn-taking, listening sounds like “mm-hmm”, and voices chosen by accent, age and gender. Text chat when voice isn't possible.
Emotion and rapport
Each persona's emotional state and trust shift with what the trainee says, so outcomes depend on skill, not luck.
Scenarios without engineers
Write the persona, add knowledge files and example conversations, and a simulation is ready in about a minute.
Structured programs
Chain steps such as Learn, Train, Debate, Panel, Interview and Assess into a program, with ordering and retakes.
Assessment and gradebook
Yes/No, grade and levels rubrics, rubric import from PDF, written feedback, and weighted totals ready for your LMS.
Integrations
LTI 1.3 with deep linking and course rosters, plus API keys for your own tools.
Conversation analytics
Talk-time share, question–answer match, follow-up questions, minutes and completions for every cohort.
Access control
Restricted or link-based sharing, participant groups, and an admin console for each organisation.
Researchers who have scaled learning technology before.
Founded in the Advanced Learning Technologies Lab at the University of Toronto. Our cognitive-architecture work continues within the Vector Institute's DaRMoD accelerator.
Dr. Steve Joordens
Chief Executive Officer
Full-time Professor at UTSC, co-founder of peerScholar, 3M National Teaching Fellow, and a global leader in psychology and assessment research.
Dr. Brian Harrington
Chief Research Officer
Professor at UTSC. Computer scientist specializing in natural language processing and AI, co-directing TalkLab at UTSC.
Robin Walker
Chief Technology Officer
CTO of peerScholar, with a deep track record building and overseeing educational technology deployments for 100k+ students in 10 countries.
Sean Ramsay
Chief Operating Officer
30+ year career in tech. Founder, veteran digital media innovator, start-up mentor and philanthropist, with 7 patents. Inventor of the first consumer 360º camera.
Supported by a team of 14 advisors, partner professionals and engineers.
Meet the team