The most important input in any room is usually the hardest one to protect. A pastor steps onto a platform while the band is still playing. A presenter closes a session over a musical tag. A vocalist works the front of a stage, a few feet from a loudspeaker that is already running hot.
In every one of those moments, the microphone is doing exactly what we asked it to do, and exactly what we wish it wouldn't: Capture the person and everything around them. Stage bleed, PA spill, room reflections, crowd noise, the scrape of a headset against a collar. The engineer needs more level. The room says no.
For most of my career as a front-of-house engineer, the answer to that problem was a collection of compromises. EQ notching to pull down the frequencies most likely to ring. Gating and expansion to shut the microphone when no one was talking. Careful mic selection, disciplined stage volume, and a whole lot of fader-riding.
Each of those tools works a little bit, and each of them costs something, usually a piece of the voice you were trying to protect in the first place. Our technology, Source Intelligence, changes that, because we approached the problem differently.
The Threshold Problem
To understand why this is different, it helps to look at how conventional noise control actually behaves. The majority of voice noise reduction tools in live sound are built around a threshold. When the signal sits above the threshold, it passes; when it falls below, it is reduced.
That logic works cleanly when the only thing in front of the microphone is the speaker or singer. It breaks down the moment the environment gets loud, because crowd wash or a drum kit next to a singer crosses the same threshold the voice does. When the voice is open, the noise rides through with it. An engineer ends up chasing the threshold all night: one setting for the quiet verse, another for the loud chorus, another for the moment the room erupts.
That is the limitation Source Intelligence is built to address. Rather than deciding whether to open or close a channel, machine learning source separation identifies the voice within the signal and reduces what is not voice—and it keeps reducing the non-voice content while the voice is active. That is the meaningful distinction. The drums next to the vocal mic stay suppressed even as the singer is singing. There is no threshold to set and nothing to chase as the show evolves.
This is not the same thing as feedback suppression, and it is worth keeping the terms straight for spec purposes. Feedback tools target the ring. Source separation targets the entire field of unwanted sound around the voice: bleed, spill, ambience, handling noise, and reflections. The reduction in feedback risk is a consequence of a cleaner source, not the mechanism itself.
What 'Cleaner' Buys an Integrator
The number people latch onto with Source Intelligence is 40 dB … of non-voice content removed from a speech or vocal microphone. I would gently steer integrators away from fixating on that figure, because the ceiling is rarely the point.
In a real room, the win is often a handful of extra decibels of gain before feedback on a spoken word channel—the difference between a presenter sitting comfortably in the system and an operator feeling like the mic is living on the edge.
At North Point Community Church in Atlanta, the team described needing only a little more margin on the speech bus, and finding that the extra dB or two was exactly what closed the gap. Modest on paper, decisive in the room.
The second benefit matters more for installed systems than people expect: what happens when you clean up several open microphones at once. Four vocal mics in front of a live drum kit each carry some bleed, and collectively that bleed reshapes the mix. Remove it from all four and the result is a markedly cleaner mix with the tonal character of the voices intact, but more importantly, notably higher intelligibility. That transparency precision and clarity is the property installers should test for first.
For show-critical and install work, the platform question is as important as the audio. Source Intelligence runs as a licensed application on a dedicated, closed audio processor rather than on a general purpose computer exposed to the internet, automatic updates, and software conflicts.
For an integrator that has to warranty uptime, that distinction is a practical maintenance and reliability decision. It also means the same processing can be deployed across mono, stereo, distributed, or immersive configurations. It's not tied to any one room geometry.
What It Is Not
Latency is the other gating spec. Machine learning inference has historically traded latency for quality, which is acceptable in a studio and disqualifying on a live stage, where the tolerance is measured in single-digit milliseconds before performers perceive a disconnect. Total input-to-output latency for Source Intelligence, including conversion and processing, is under 8.5 ms. We treated that figure as a fixed ceiling: Every improvement to the model had to stay under it before we would ship it.
Here is the honest part. We spent more than 18 months testing this in the field, including on major touring productions for Harry Styles, Bruno Mars, The Weeknd, and Rosalía, before we called it a product. Even so, Source Intelligence works alongside good system design, sound mic technique, appropriate stage volume, trained operators, and acoustic judgment, rather than standing in for any of them.
It is not a rescue for a room that was never tuned. It is a new kind of margin in the moments where margin has always been hardest to find: when someone needs to be heard clearly while the room around them refuses to be quiet.
For most events, the voice carries the message—the information, the emotion, sometimes the entire reason the room is full. Source separation, whatever badge it wears, is finally giving us a way to protect that source without trading away the thing that makes it worth protecting.
