The rapid advances of AI technologies have enabled new learning and development (L&D) and organizational development (OD) applications in which a learner talks to an AI, rather than typing into a chat. This opens up new classes of learning solutions for L&D.
One of the most popular uses of voice-to-voice in L&D is for AI role play. Role play is a useful tool for L&D and current AI models are remarkably good at it when properly set up.
Where AI role play helps
The most common uses are for sales, customer service and leadership training. Key to success is to design a variety of custom role play scenarios that closely map to what the learners will face on the job. Many companies already have well-designed rubrics and scenarios for sales enablement which can be turned into AI role play. With AI role play L&D can deploy solutions which could speed up time-to-competency for a broad variety of training needs such as: new employee training, new manager training, new customer service personnel training, executive training, new sales rep training, new product training for sales reps, new nurse training, new police officer training.
With AI role play sales reps can practice pitching a new product until they know what to say and how to say it, before getting in front of a real prospect. They can practice defending value when a customer has a cheaper quote on their desk, or handling the polite rejection when working the trade show booth, reopening a stalled deal, rehearsing a check-in when a loyal account starts drifting, and more.
AI role play is equally useful for giving young financial advisors a chance to practice helping a client who is panicking about a stock market crash and wants to pull their money out, or for new bank tellers to practice responding to a customer who needs the money on the check they deposited yesterday but it has a hold on it.
In leadership training, AI role play enables new managers to practice difficult conversations in a private safe space with a simulated character. They can practice holding a project manager who blames others for a slipped deadline accountable, or practice having an employee career conversation, or delivering negative feedback productively.
Aside from AI role play, other uses for voice-first AI in L&D include AI coach systems. They can provide personalized support in between coaching sessions, or guide a learner through a reflective discussion around a goal.
Why it suddenly works
I spent fifteen years building speech recognition middleware and applications, from Bell-Northern Research to SpeechWorks and Nuance Communications. Every voice application of that era worked in steps: recognize the words, decide what to do, then play or synthesize a reply.
The first voice applications built on large language models (LLMs) relied on cascaded systems. A user talks and their speech is transcribed, sent to an LLM, and the output is run through a text-to-speech engine so the user can hear the response. Even well-built systems typically leave a second between turns, which is significantly more than the roughly 200 ms gap in a typical conversation. Also, all the additional information that humans use to communicate during a conversation such as tone, pacing, hesitation, and inflection is lost during transcription leaving the AI with only the user’s words to work with.
The new speech systems are native audio. They have been trained natively on audio, as well as text, enabling them to take user speech directly as input, and output speech directly as audio. There is no need for transcription or text-to-speech so these systems respond much more closely to how a human would, and ‘react’ both to what was said and how it was said. Native audio systems allow an AI application to manage, and perhaps sooner than we think, even facilitate, increasingly sophisticated sessions with learners.
What to watch
As AI-powered speech systems gain traction in L&D, it’s worth keeping an eye on evolving privacy laws and regulations around the use of AI. The EU bans AI systems from inferring emotions from biometric data, including voice, at work or in education, except for safety or medical reasons. Quebec’s Law 25 only allows personal information to be processed outside the province after a privacy impact assessment shows it will be adequately protected, and under a written agreement.
The voice practice system I built doesn’t infer, label or score users’ emotions from their voice. CoachingOurselves works with its clients on what Law 25 requires, and on their own IT policies for cloud-based AI apps.
Try a live native audio AI role play demo right now.
I designed and built the voice-to-voice system behind this demo. The voice practice is sold through CoachingOurselves, which builds custom voice-first learning solutions for L&D and OD, drawing on almost twenty years of peer learning design.
What AI can’t do
Though voice AI solutions are useful, they cannot build the human connection and what Henry Mintzberg calls communityship, vital for successful organizations. Dialogue, and the guided discussions in CoachingOurselves peer learning groups are a simple, scalable human approach to leadership and organizational development.