Barge-in is the most expensive feature users assume is free

Voice UXAudioCost
Updated 2026-09-08

Letting a user interrupt a talking device mid-sentence, by simply speaking, is a systems problem with no cheap and reliable answer at consumer hardware cost. We shipped a dedicated noise-reduction IC and an algorithm for exactly this, tuned it for weeks, and that is still the honest conclusion.

What we were trying to do

Barge-in looks trivial. You start talking, the device stops and listens. It is basic conversational courtesy, and every user assumes a voice AI should behave this way. We treated it as a detail, budgeted hardware for it, and found a hole.

Where it actually breaks

Testing in our own office, a colleague was on a phone call nearby. The device took his voice as an interrupt and dropped the context of the conversation it was having with me. Not a distant sound, not a loud one, just another adult talking at a normal level a few metres away.

The root cause is that barge-in is not a loudness question. To decide correctly, the device has to know three things at once: who is speaking, where they are, and whether what they said was addressed to this device. Each is hard alone. All three, on a cost-controlled consumer board, is close to unsolvable.

Four approaches, and how each one fails

Distance

A child carries the toy around, puts it on a table, holds it up. Distance between mouth and microphone is never stable, so any threshold built on it drifts constantly. This route is closed before you start.

Volume threshold

Ambient sound overlaps the user's own speech in level. A television, a nearby conversation, street noise through a window all land in the same band as the person you care about. The same measured level can be the owner speaking or the TV, and nothing in amplitude alone separates them.

Voiceprint

Technically the best answer. Lock to the enrolled user and ignore everyone else. It needs compute and licensing that a consumer toy BOM does not carry, so it is priced out rather than ruled out.

Local noise-reduction IC

This is what we actually shipped. It helps, and in a reasonably quiet indoor room the behaviour is acceptable. In a noisy open environment false triggers still happen. We did not solve the problem, we moved the failure rate.

Why phone assistants get away with it

A phone assistant handles barge-in well because it has a microphone array, a high-performance application processor, and cloud compute behind it. Move those into a cost-controlled toy and each one breaks something. The array adds size and cost. The processor blows the BOM. Cloud round trips add latency and a network dependency in a product that has to work when the Wi-Fi is bad.

This is not a question of whose algorithm is cleverer. It is a different constraint set.

What we shipped

We tightened the trigger boundary to cut false positives, accepting that the device will occasionally miss a genuine interrupt rather than constantly react to the room. It responds reliably when the user says its name or speaks clearly toward it. Zero false triggers in an open, noisy environment is not achievable at this cost today, and we would rather say so than let a channel discover it after launch.

The general lesson

In AI hardware, the interactions users consider obvious are frequently the most expensive to engineer. Barge-in is one. If you are scoping a voice product, price this feature explicitly instead of assuming it comes with the voice stack, and decide early which direction you want the errors to fall.

Want to try the hardware behind these notes?

Six AI hardware platforms are in mass production. Tell us your market and channel and we send an evaluation sample.

Request a sample