Field sales reps for pharma and healthcare accounts spend most of a working day between appointments: driving between clinics, walking a hospital corridor, waiting in a lobby. That’s dead time for training, because most rep-enablement tools assume a screen, a keyboard, and a quiet desk. We built an AI coach that drills pitch lines, objection handling, and compliance phrasing entirely by ear. A rep runs a full practice session from the car without touching the phone.
The challenge
A voice-only coach has to work in the environment reps actually train in: a moving car cabin, a hospital hallway with overhead announcements, a conference floor between meetings. Most voice AI demos are recorded in a silent room and fall apart the moment real background noise enters. The system mishears the rep, cuts them off mid-sentence, or leaves the microphone open picking up ambient noise. For a tool that lives entirely in the ear, that failure is the whole product failing.
How we handled the noise
Three design choices made this hold up outside a quiet room.
Capture-level suppression, on by default. Every recording session opens the microphone with echo cancellation and noise suppression enabled at the browser audio layer, and pins the sample rate to what the speech engine expects. The raw signal is cleaned up before recognition runs, not after.
Endpointing by energy, not by a clock. A fixed-duration timer is the first thing that breaks in a noisy cabin. It cuts off a rep mid-line, or leaves the mic open into road noise waiting for silence that never arrives. We replaced the timer with RMS energy-threshold detection. The system watches the actual loudness of the incoming audio and ends the turn when speech stops, not when a clock runs out, so it adapts to whatever noise floor the rep happens to be in.
Scripted turns, scored against a known line. The coach always knows which pitch line or objection response it just asked for, so recognition never has to guess at arbitrary words in a noisy room. It scores a specific, expected phrase against what came back. Bounded recognition like this tolerates far more background noise than free-form transcription, because the system already knows the answer it’s checking for.
The fix that mattered most
Almost every noise-related failure traced back to timing, not intelligence: a fixed timer guessing wrong about when the rep was done. Swapping the clock for an energy-based endpoint fixed more sessions than any change to the recognition model itself.
What a session actually sounds like
Every step in the exchange is built to work with no screen involved, because a rep driving between clinics or walking a hospital corridor doesn’t have a free hand for one.
A short ascending two-tone cue plays in full before the microphone opens, so the rep always knows exactly when to start talking. The app then waits 400 milliseconds before recording starts, long enough for the coach’s own voice to stop echoing off the car cabin or hallway walls, so that leftover echo never gets picked up as part of the rep’s answer. Recording ends on 1.5 seconds of measured silence, or a 10-second hard cap, whichever comes first, and a descending two-tone cue confirms the mic has closed. The rep’s own recording plays back once before the coach responds, so a rep with one earbud in and both hands on the wheel still knows their answer was captured, with no screen to check.
The result
A rep runs a full coaching session in a car or a hallway, without opening an app or looking at a screen, in the same noise the job actually happens in.