Smooth Feedback Hunter

Feedback suppressor plugin for live sound.
macOS (VST3, Audio Unit) and Windows (VST3), zero added latency.

What it does

Catches the ring, not the voice

Trained on measured rooms, from a small wooden room to a stone-walled manor hall, it takes gain off the feedback and not off the performer: a sung or spoken note is left alone, a ring that builds up out of nothing is cut, and when the ring is gone the cut is gone. A ring the room has already shown once is caught faster the next time.

Zero added latency, nothing held back

The audio path adds no delay at all — nothing to compensate for, nothing that drifts out of sync with the rest of the desk, and no pause gate. On our own stage this gives clean speech at 50 cm from the microphone, next to the speaker.

Works on a group bus

Put it across a whole group — a cast on lavaliers, several singers, a choir, voice and piano — with nothing to switch: it recognises what it is hearing by itself. A group costs more level than a solo voice: in this build six singers on one bus come through 2.6 dB down and a choir 2.9 to 3.1 dB down (the previous public build 0.28.58: 2.5 and 3.5 to 4.3 dB), but voice and piano 5.2 dB down (0.28.58: 2.7 dB). Stereo buses keep their image.

New: the ring stops when the sound stops

When a knock or a hit ends, the ring it leaves in the room dies away within half a second at the ring point: a quarter to half a second after the hit it is 37 dB quieter than with the previous public build 0.28.58. The first words after a long pause come back within 1.5 dB of full level, and long sung notes are rarely cut off suddenly.

Learned noise print

Hiss from a cheap microphone or a fan at the back of the hall is learned in the quiet moments and taken out of the output, with the NOISE knob setting how much — still at zero latency. Hold the knob for a second to teach it the room's noise yourself, click it to go back to automatic.

+19 dB more stable gain

Median added stable gain with speech across eight measured rooms, from +13 dB to +21 dB. Howl held down at +8 dB over the ring point in 79 of 80 closed loops and at +12 dB in 79 of 80. Inside a ringing loop the voice comes through nearly 3 dB louder than with 0.28.58. Measured in a closed feedback loop driven by real room impulse responses, loop gain raised until the system rings, bypassed versus engaged.

Measured

Where 0.28.77 stands on our own test bench (re-measured 5 November 2026 on the released binary; the previous public build 0.28.58 is given in brackets).
All figures come from the released plugin binary, not from a lab version. Your room, your microphones and your speaker placement will give you different numbers — these are measurements, not a guarantee.

+19.0 dB median added stable gain with speech across eight measured rooms (range +13 to +21 dB; 0.28.58: +20.0 dB, range +15 to +24 — three rooms now higher and five lower). Closed feedback loop driven by measured room impulse responses; loop gain raised until the system rings, bypassed versus engaged.
Howl held down at +8 dB over the ring point in 79 of 80 loops, in 79 of 80 at +12 dB and in 77 of 80 at +16 dB (eight rooms, two voices, five loop delays; 0.28.58: 80, 80 and 76). On the wider 11-room bench: 107, 108 and 107 of 110 (0.28.58: 110, 109 and 103; repeated runs of 0.28.58 moved its +16 dB count between 101 and 109).
Rings taken out: on a separate bench of 420 closed loops with 1 668 known ring modes, 0.28.77 removes 78 % of them and leaves 1 % ringing (0.28.58: 83 % and 2 %, Defeedback AI 1.1.4: 36 % and 22 %): 78 % of the rings below 2 kHz (0.28.58: 84 %, Defeedback AI: 41 %), 96 % of those between 2 and 4 kHz (0.28.58: 78 %, Defeedback AI: 26 %) and 76 % of those between 4 and 9 kHz (0.28.58: 78 %, Defeedback AI: 11 %). Repeated runs of the same version move these figures by up to 8 points.
Speech level: −1.2 dB median on a clean speech recording in the open path; the quietest tenth of the frames sits 3.5 dB down (0.28.58: −0.5 and 2.9 dB). 13.2 % of the time–frequency tiles lose more than 6 dB (0.28.58: 3.0 %). Inside a closed loop at +8 dB over the ring point the voice comes out 5.5 dB lower than unprocessed (0.28.58: 8.2 dB), with 5.4 dB of residual ring on it (0.28.58: 0.6 dB). In a hard, reverberant room about 3 dB of the room's own ringing is also taken off the voice on a microphone recording from a stone-walled hall (measured with 0.28.13), the same as Defeedback AI on the same recording. On an opera recording 11.4 % of the time–frequency tiles lose more than 6 dB and the median level is 1.6 dB down (0.28.58: 8.3 % and 1.7 dB) — among the weakest figures on this bench, with groups on one bus (below).
Groups and instruments on one bus (open path, level over active frames; 0.28.58 / Defeedback AI at full strength in brackets): six singers −2.6 dB (−2.5 / −5.5), a choir-like mix built from real singers who were kept out of training −2.9 to −3.1 dB (−3.5 to −4.3 / −4.2 to −5.5), voice and piano −5.2 dB (−2.7 / −3.5). Voice and piano lose clearly more than with 0.28.58. Choir and piano together remain hard: −14.3 dB (−11.4 / −5.3). In 44 closed loops with voice and piano or a unison choir as the source, 37 of 44 stay free of a strong ring at +8 dB over the ring point and 32 of 44 at +12 dB (0.28.58: 38 and 41; Defeedback AI: 11 and 4).
After a pause: after 5 to 60 seconds of silence or quiet room noise the first words come back within 1.5 dB of their full level (0.28.58: 2.1 dB on the same test).
Ring after the sound stops (knock phrases in closed loops, eight rooms, ring level relative to the hit; 0.28.58 / Defeedback AI in brackets): at the ring point 0.1–0.25 s after the last hit −56 dB (−40 / −61), 0.25–0.5 s −85 dB (−49 / −74), 0.5–1 s −105 dB (−64 / −75). At +6 dB over the ring point 0.25–0.5 s −47 dB (−34 / −68) and 0.5–1 s −93 dB (−46 / −72). Defeedback AI is still quicker in the first quarter of a second and at high gain.
Latency and CPU: zero added latency; about 9.5 % of one CPU core per instance at 96 kHz with a 64-sample buffer, all of its threads counted (Apple M4 Pro; in the same run 0.28.58: 3.3 % and Defeedback AI 1.1.4: 12.5 % — this build needs nearly three times the CPU of 0.28.58 and less than Defeedback AI), noise removal included. Output peak guaranteed below −1 dBFS.

Compared

The same bench, five plug-ins: Smooth Feedback Hunter 0.28.77 (re-measured 5 November 2026), Defeedback AI 1.1.4, CALM 2 2.4.6, Feedback Guardian AI 1.0.0 and VoiceISO 1.0.6, all as released binaries. Closed feedback loops with measured room impulse responses (8 rooms × 2 voices × 5 loop delays), speech and singing in the open path, CPU at 96 kHz on an Apple M4 Pro with all of each plug-in's threads counted, re-measured for all five on 5 November 2026 (CALM 2 is an Intel build, run under Rosetta). Product names belong to their owners; the figures are our measurements and we will show the method and logs on request.

System requirements

macOS	Available macOS 11.0 or newer, universal binary (Apple Silicon and Intel). Signed with a Developer ID and notarized by Apple — the installer opens without extra steps. VST3 · Audio Unit
Windows	Available Windows 10 or 11, 64-bit. Signed installer (SFH Digital OÜ, Microsoft Trusted Signing); one licence covers both platforms, one computer at a time. VST3
Licence	One licence runs on one computer at a time. Activation and deactivation happen in the plugin and need an internet connection. The plugin confirms the licence with our server every time it loads while online; without internet it keeps working for 21 days after the last confirmation. Moving to another computer: DEACTIVATE in the plugin, then activate on the new one with the same key. Broken or lost computer: write to info@sfhdigital.com and we release the licence for you. Fair use: up to three different computers per licence in any 30 days.

Try it first

The download is the full plugin. Without a licence it runs in demo mode: everything works, but the audio mutes for about seven seconds every two minutes. Enter a key and the muting stops. Nothing else is held back, so you can judge it on your own rig before paying.

After you buy

Your licence key is emailed to the address you pay with, usually within a minute. Open the plugin, click + in the LICENSE panel, enter your name and the key and click ACTIVATE (internet needed). One computer at a time: DEACTIVATE in the plugin before activating elsewhere — the plugin waits for the server to confirm the release. Lost, broken or offline computer? Write to info@sfhdigital.com from the purchase address.

Questions

What is a feedback suppressor?
A processor that stops acoustic feedback — the howl or ring that builds up when a microphone picks up its own loudspeakers — so you can run more gain before feedback. Classic feedback suppressors and feedback eliminators drop narrow notch filters on frequencies that have already started to howl. Smooth Feedback Hunter works on the whole spectrum, band by band, and tells the room's ringing apart from the performer, so the voice keeps its tone.

Where do I put it?
On the insert of the microphone channel that feeds the speakers, in any host that loads VST3 or Audio Unit plug-ins. The LIVE path adds no latency, so it can sit on a vocal or speech channel without putting it out of time with the rest of the desk.

Does it replace ringing out the room?
No. Good speaker and microphone placement and a sensible EQ come first. The hunter catches what still rings when a microphone moves, a presenter walks in front of a speaker or the level goes up during the show.

Will it colour speech or singing?
On our bench clean speech loses 1.2 dB median in the open path: most of the gain comes off the room's ringing, not off the voice. Between phrases the ringing is pulled down, so the quietest tenth of the frames sits about 3.5 dB down. On an opera recording 11.4 % of the time–frequency tiles lose more than 6 dB. In a hard, reverberant room about 3 dB of the room's own ringing is also taken off the voice, on a microphone recording from a stone-walled hall.

When can I get it?
Now: the 0.28.77 beta for macOS and Windows is free to download above and runs in demo mode until you enter a key. Beta testers who want to report back can use the beta programme page.