AI vs. Feedback: This Fight Started Long Before AI

Feedback has been tormenting live audio engineers for roughly as long as we've had microphones pointed anywhere near loudspeakers. Move the mic. Move the speaker. Ring out the room. Find the frequency. Pull it down. Open another lav. Pull something else down. Ask the presenter to please stop walking directly in front of the PA. Repeat until doors.

So when products like Alpha Labs' De-Feedback began appearing with claims of AI-assisted feedback reduction, it was tempting to view them as an entirely new solution to an ancient problem. The technology may be new, but the goal isn't. Engineers and manufacturers have spent decades trying to answer essentially the same question: How do we make the microphone hear more of what we want and less of everything else?

What's changing now is how the processor answers that question.

Traditional feedback control starts with physics: microphone and loudspeaker placement, polar pattern, distance, system tuning, EQ and gain structure.

When those options aren't enough, automatic feedback suppressors identify troublesome frequencies and insert narrow filters. Waves' X-FDBK, for example, automated a process engineers had already been performing manually for years. It works, but every notch still changes the frequency response of the microphone.

Anyone who has finished ringing out a difficult ballroom and stared at the resulting EQ curve like it was the elevation map of the Rocky Mountains knows the tradeoff.

Another approach is reducing the microphone's contribution to the system when the wanted source isn't present. The Rupert Neve Designs 5045 Primary Source Enhancer became one of the best-known examples. The all-analog processor detects when someone is speaking or singing and reduces gain when they stop. Rupert Neve Designs says it can provide up to 20 dB of additional gain before feedback under the right conditions.

Waves later brought a similar idea into software with the Primary Source Expander, or PSE.

These are clever tools, but they're largely making decisions based on signal behavior. The newest generation is trying to answer something harder:

Is this sound actually the person?

CEDAR belongs in this story too. Its DNS1000 Dialogue Noise Suppressor, introduced in 2000, brought near-zero-latency noise suppression into broadcast and other real-time environments. The later DNS 8 Live pushed the concept further for live sound and broadcast.

That history matters because AI didn't suddenly arrive backstage in 2026 carrying a laptop and asking for the Wi-Fi password. Noise suppression, adaptive processing and increasingly intelligent decisions about microphone content have been developing for decades.

Alpha Labs released De-Feedback V1 in November 2024. Instead of simply waiting for a frequency to ring or attenuating a microphone between phrases, the processor uses a trained AI model to reduce feedback, room reverberation and background noise while attempting to preserve the vocal itself.

De-Feedback is extremely CPU intensive, which makes support on platforms such as Waves SuperRack LiveBox and Fourier Audio transform.engine especially important. Those systems helped move the technology from an interesting experiment into something engineers could realistically deploy on a major show.

But perhaps the most telling part of the De-Feedback story isn't the specification sheet. It's where the thing has already been used.

Robert Scovill used De-Feedback during the 2025 Rock & Roll Hall of Fame induction ceremony and later with Fito Páez and Kenny Chesney. His description of the Rock Hall results: “mind-bending.”

Iron Maiden FOH engineer Ken “Pooch” Van Druten has also incorporated it into Bruce Dickinson's vocal chain, particularly useful on a stage where Dickinson can wind up frighteningly close to the PA. Pooch called the improvement a “game changer.”

The use cases keep piling up. De-Feedback has also appeared on Mariah Carey's Christmas tour, referee microphones during an NFL playoff game, Broadway's Every Brilliant Thing, and the Survivor 50 Live Finale, where FOH engineer Andrew Fletcher reported having as many as 24 lavaliers open simultaneously.

That's a pretty rapid progression for a product that went on sale in late 2024.

Then L-Acoustics made the story considerably bigger.

Introduced in 2026, Source Intelligence runs on the L-ISA Processor II Core and uses machine-learning-based source separation to distinguish human voice from crowd noise, instrument bleed, PA spill and room reflections.

This isn't another VST running on a third-party server. It's source separation becoming part of the reinforcement ecosystem itself.

L-Acoustics currently specifies four instances, latency below 8.5 ms and as much as 40 dB of non-voice signal reduction.

With a traditional feedback suppressor, we're manipulating the system around the source. With an expander or 5045, we're controlling when the microphone contributes gain. With Source Intelligence and De-Feedback, the processor is attempting to determine which part of the microphone signal is the source.

The fight against feedback is becoming source separation.

The market is filling up quickly.

In late August 2026, CEDAR released Voxis, a low-latency neural-network voice isolator specifically optimized for live environments. With latency below 10 ms, CEDAR is targeting live spoken word, singing, FOH, conferences, live sports and broadcast—bringing the company that launched the DNS1000 more than 25 years ago directly into the current neural source-isolation conversation.

Clear Voice Live is another entrant, combining denoise, dereverb, debleed and feedback prevention in a real-time AI voice processor. Its current release offers both a zero-latency mode for paths feeding stage monitors and a 127-sample, 2.65 ms low-latency mode for other applications.

VoiceISO takes a similar approach, using a deep neural network trained on vocal, reverberation and noise data to identify direct vocal sound and reduce reflections and unwanted material in live reinforcement. Like De-Feedback, its algorithm itself adds no latency, while actual system latency depends on the hardware and audio interface around it.

Smaller developers are experimenting in the same territory as well. Everworks currently offers products including CALM 2 and other source-aware cleanup tools, another sign that real-time vocal isolation is becoming its own category rather than a one-off curiosity.

For most of live audio history, feedback management has asked: How loud is the signal? What frequency is ringing? Is the microphone being used?

These new processors ask something different:

What am I listening to?

That's a much bigger development than simply finding the feedback frequency faster.

None of this replaces good system design. Microphone placement, loudspeaker deployment, gain structure and system tuning still matter, and AI processors bring their own concerns around latency, processing power, artifacts and reliability.

Physics has not received a firmware update.

AI just gives engineers another tool for dealing with it.

It's becoming easy to imagine source isolation joining EQ, compression and gating as a normal part of the live channel strip: voice isolation, bleed reduction, dereverberation and feedback protection.

We've spent decades chasing feedback frequencies around the room. The new approach asks a simpler question:

What if we could keep more of the person—and less of the room?

For anyone who has spent an afternoon trying to get twelve lavs loud enough in a ballroom, that's worth paying attention to.

-CTC Staff