AI De-Feedback/Anti-Reverb Demo + Full Explanation

[@basspuppy133]

These are the sort of things AI should absolutely be applied for, Very cool demo.

[@mashzmash]

Please license this to Waves so it can run natively on LV1 using SoundGrid :)

[@devinlsheets_alphasound]

Email them about it. They could choose to do this tomorrow if they wanted to.

[@zazyczech]

Well, this is basically live mode of Clarity isn't it?

[@devinlsheets_alphasound]

@@zazyczech no, because clarity doesn’t prevent feedback.

[@nielsott]

Waves have their SuperRack Livebox that can run VST3. Which btw. also features Dante. For just $6.999 you can get there. Me & myself, I use a Mac Mini M1 with a RME Digiface Dante and Superrack Performer and I also get below 5msec latency with that system with plugins that report zero latency to the host. I don't know however if this system would survive this advertised Anti-Reverb… also you can't upgrade RAM on Macs, so this might be limiting, too.

[@devinlsheets_alphasound]

@@nielsott We have a Livebox, works great! One instance will work with the lowest latency settings at 48k. Two or more instances require increases in buffer size, until eight instances will run with 96 sample buffer size and that's the largest buffer size we'd recommend for live sound.

[@bob4analog]

It's just amazing how this can work in real time with no latency! This is like femto second technology!

[@WeaselFace]

I'm skeptical but open minded. I'll give the vst a trial tomorrow in a very reverberate, live sounding concert hall i work in quite regularly. I'll run it on a lectern and I'll do an A/B test between my experience with a graphic and parametric, and some good gain structure, then i will try a clean signal with just the VST and see how things compare. Seems like the right use-case for this plugin, am i right?

[@devinlsheets_alphasound]

This is exactly what we designed it for. Let us know how it goes

[@WeaselFace]

@@devinlsheets_alphasound Tested today in an empty 1000PAX concert hall with a single presenter at lectern and it did pretty well...however, I can't give it the full sell just yet as the "audio in/out trial" version means I cant put it to test on an actual show. I have a bunch of corporate lectern shows coming up as well as a few rock concerts I'd love to throw this vst on lead vocals for. I want to see how it deals with loud crowds and loud stages, but without being able to test it live I cant justify such a high ticket price for the license. Is there, like, a 30 day trial version instead of the "audio in-out" that I can use on a few shows?

[@aaronabeytia]

This is an excellent concept; excited to see what additional developments will bring. There's definitely a place for it, and further refinements can help it grow to the point that it's widely used!

[@DennisRudiNielsen]

Will it isolate a live vocal by removing just the reverb and feedback or can it also remove cymbal bleed and other stagenoise that is not a vocal?

[@devinlsheets_alphasound]

at this time, it does not remove much else besides feedback and reverb. perhaps future versions will remove more

[@elipto8836]

there is a plugin for drum and cymbal bleed already. Drum gate debuted

[@nTOXICATEDAV]

@@devinlsheets_alphasoundHello. Sound engineer here. Please add cymbal bleed removal to it also. This will be even more huge for drummers who sing (and are also main singers). Also please contact Behringer, so they can add this to their Behringer and Midas consoles/mixers.

[@wietzejohanneskrikke1910]

Yes, but it does degrade the audio. It works better than a standard gate, but still...

[@devinlsheets_alphasound]

@@wietzejohanneskrikke1910 in “live” listening tests, on a real sound system, we can’t hear any audio degradation. In fact, the overall quality of the sound goes up significantly because of what the plugin is doing

[@ffE123IJK]

This is impressive. It's definitely going to help.

[@mehi8145]

Would the AI model work even better if it would get the IR of the venue?

[@devinlsheets_alphasound]

Possibly, but 1. the difference in a live sound setting may be so slight that it’s not worth the trouble and 2. a sufficiently trained general intelligence should be able to handle all rooms well enough I think we’re pretty close to that already based on having used it for about a year now myself, and from the reports of others. Still room for improvement of course but I think that improvement is mostly to be found on the current path which would not involve a “learn” condition on site.

[@Jaburu]

this technology should also apply to the playback of a speaker in a reverberant room?

[@PrincipalAudio]

Wow, amazing job! Well impressed!

[@FREEZYREVS]

Would love to see the application of this on instruments in a concert setting !

[@devinlsheets_alphasound]

For now it only works with vocals, but future versions may handle instruments

[@Ozoingo2]

Its really promising, i found it increased my gain before feedback by a few db standing right in front of a speaker. It feels like the (reaction time?) has a long way to go. At high gains (probably 3db above a good ring out) it seemed to have high frequency artifacts in the vocal. Still, once its stable and signed on mac i can see it being a really nice tool. Thank you for your work, hats off to ya

[@instrumentenfreak]

This is impressive. Will it be possible without any artifacts?

[@devinlsheets_alphasound]

Yes, if we sell enough of the current version to become profitable enough to ramp up the AI computer and data sets significantly

[@Telecraster]

What are the silicon restrictions of what this could run on in the future other than x86. Applicable to SoC/appliance-based applications potentially too?

[@Puppetmaster2005]

Have you guys tried this for conferencing applications? As in.... together with some ceiling microphones such as the Shure MXA920 and using it for voice lift. The main challenge of voice lift is the fact that we have to EQ away the feedback, thus destroying most of the tonal quality of the pickup in the process. I'm wondering if someone somewhere out there has tried running the audio from the ceiling mic into your plugin, then back out into the room speakers? That would be a really cool application of this AI de-feedback imo.

[@devinlsheets_alphasound]

We haven't tried that specifically but it would probably help. It isn't designed specifically for that, but someone should try it and tell us how it works!

[@bob4analog]

I wonder if there's a break point, where it will feedback when you get too close to the speaker with the mic. Or is it absolutely NO feedback. That would be interesting to get right up to the speaker with the mic and then speak into the mic.. with no feedback!?! Amazing!

[@devinlsheets_alphasound]

It won’t feedback if you aren’t talking into it. If you’re talking into it, and you’re too close to the speaker, and it’s turned up too much, you can defeat the current algorithm. Future versions will be much harder to defeat. But as is, the current version adds 6-10dB more gain out of the gate. Not bad.

[@jxnxsdev]

Does this work on a compiled mix of multiple microphones with people talking over each other, or does every microphone need a seperate instance of the plugin? Also what happens when music plays while people are talking? Will the ai also be able to filter out music feedback?

[@devinlsheets_alphasound]

It can work with two or three people talking over each other at once, but it is best to only have one person at a time. It won't filter out music in the background, yet.

[@StatiXonfirE]

I have no idea what I’m talking about… but could it be put on an output Channel?

[@devinlsheets_alphasound]

You could... as long as the only thing going through that output are singular human vocals, and not too many at one time. Although, some people are trying it for choir with success, but we don't "officially" recommend this yet.

[@StatiXonfirE]

@ so would it work on like a vocal/mc bus/matrix if that was run through a direct out and then back in on a channel?

[@devinlsheets_alphasound]

@@StatiXonfirE sure

[@MFKitten]

Ideally you guys would find a way to get this running on linux, and figure out a dedicated hardware+software package that lets you build a PC and install a distro that literally only starts up, launches the defeedback AI and the audio interface drivers, and nothing else. With no need for any user input. The absolute winner would be a rackmounted unit with a power on/off button and XLR in and out.

[@devinlsheets_alphasound]

That would be great. Want to invest in the company and make it happen? Btw the hardware units we sell on the website… once you activate them, they function basically like what you describe, they just aren’t rackmount units, although some people have put them into rackmount cases.

[@derekauth157]

I wonder if Rode would ever use something like this in some of their lineup

[@urbanusta]

Thanks for the update. Use of described processing removes reverberation as picked up at the position of the processed microphone and prevent exacerbation of the same via PA. However, any other listener (or an unprocessed mic) at a different location will still hear the room induced reverberation (sans PA exaggerated reverb at the processed location). Does this reflect a correct understanding? If so, the audience experience from using this in a reasonably "live" venue is greatly reduced, but still noticably present reverberation?

[@jonathanmelhuish4530]

I don't think that's right. It's not "noise cancelling", it's just removing the room response from the signal, so you land up with dry mic.

[@devinlsheets_alphasound]

If you are in a room, you will still hear the sound of the room. You just don’t hear any MORE of the sound of the room coming back through the speakers.

[@urbanusta]

​@devinlsheets_alphasound Thanks -- exactly what I expected. Great job! Would be interesting to profile this to see how many you could binpack into a Xeon class 1U server.

[@AndyPayne42]

You must get lower latency because its just a "simple" inferred formula so you aren't using a model in real time besides maybe adjusting formula at some interval. But this would mean we could offload that to a local server or even remote API so it can be thinking about the next formula concurrently. This really has me thinking, going to see if I can build this myself assuming the code/design isn't open sourced. I'll throw it up on github so others can play with it too.

[@paulforester6996]

Next you'll hear a AI voice telling you to stand away from the speaker.

[@tkhong9527]

Does it need minimum level to drive in the plugin to make it work ? Like -18dbfs ?

[@devinlsheets_alphasound]

Nope

[@dusklightmusic]

A revolutionary invention.

[@Rasenschneider]

Is AU Plugin available or planned?

[@devinlsheets_alphasound]

Yes there is a beta version for Mac available on the website

[@Rasenschneider]

@ Great, thanks!

[@SamuelHaywood17]

@@devinlsheets_alphasounddoes this include Apple silicon?

[@Ozoingo2]

​@@SamuelHaywood17 yeah I got it working on M1 relatively stable

[@AlexeyF-hb2]

Definately waiting for Waves to adopt this technology. For now- buying a PC for a single VST plugin and running audio through crappy focusrite interface with 5+ ms latency and excess AD conversions is ridiculous. This needs serious research and investments, to put the AI model in a standalone device with digital I/O and low latency, or run it as a low-latency VST3/Soundgrid plugin with reasonable processing load. A small company can't have all these resources to polish the product and make it user-friendly. You should pitch this product to Waves or at least Music Tribe! Its the only way it can become big, before these giants steal the idea and make their own AI models.

[@devinlsheets_alphasound]

Email Waves and tell them to put this on Soundgrid

[@Puppetmaster2005]

There's also a very large and untapped conferencing market out there..... Normally, large conference rooms with a lot of reflective surfaces utilizing ceiling mics, table mics, or those all-in-one conferencing bar speakerphones always run into trouble with feedback and/or too much reverb. Unfortunately, it's very common for these systems to run their audio through DSPs such as Q-sys, Extron, Biamp etc, and most of the time these DSPs don't have the tools necessary to remove all the reverb/feedback. Will it be possible for you guys to one day partner with all these big DSP brands to get your plugin into the DSP itself? That would extremely cool.

[@AndyPayne42]

Zero latency sounds too good to be true? Are the model weights open? I have a couple 4090's in my DAW I would love to use this in ardour or on its own via pipewire.

[@wietzejohanneskrikke1910]

Lol

[@AndyPayne42]

@@wietzejohanneskrikke1910 guessing you think that will be a....NO

[@sendforacar9323]

Put this in a floor unit pedal that you can plug the microphone into and out of and slap +48v on it, and this will be the greatest hardware in anyone vocal chain.

[@pkrunnerch]

I bought it a few weeks ago :-)

[@devinlsheets_alphasound]

Amazing! Do you like it??

[@pkrunnerch]

@@devinlsheets_alphasound So far I like it 😊 I usually do a lot of livestreams at different locations where we have to be ready quickly. What I haven't really been able to test yet is how it behaves when I have several different speakers on one channel. What is your experience with this? If the speakers speak one after the other, this is not a problem, but if it is in a discussion round with e.g. 5 speakers? So far I have assigned this to each channel individually. (Wing Compact and StudioRack). What is your experience with this?

[@devinlsheets_alphasound]

@@pkrunnerchhonestly it should be fine. Even if people talk over each other a little bit, it should be ok. But ideally each person gets their own plugin

[@AlexandrIZ21261]

When you activated the plugin, the sound became similar to artificially generated. This was very noticeable in the female voice in the demo. Maybe it's fine for situations where people are speaking. But for singers I think it's bad.

[@devinlsheets_alphasound]

feel free to only use it for speaking, but just fyi the sound engineers for mariah carey and the tabernacle choir at temple square both use it and say it sounds great

[@jaethomasmusic]

this needs to be in a hardware box. I still experience some feedback from my Bose system which is quite good for feedback otherwise.

[@Timkast]

🚨 RED ALERT! 🚨 Please DO NOT set a retail price for this yet!!!! PLEASE!! You can literally name your price in the broadcast audio modality!!!!!!! YOU CAN BLACK BOX THE HARDWARE SPECIFICALLY FOR COMMERCIAL BROADCAST!!! BROADCAST WILL PUT GENERATIONS OF YOUR KIDS THROUGH COLLEGE!!

[@devinlsheets_alphasound]

Ok listen up everybody. If you work in broadcast, the price is $100,000

[@Timkast]

@ 🤣🙏🏽🤣…I’m serious! If you need a Ray Krock for this I’d be honored to help. Commercial broadcasting has fundamentally changed since the pandemic; all of the sudden, kitchens, dens, offices have become acceptable remote soundstages. Huge opportunity to sell black box to network/station cluster and then license the use down to their production and remote staffs. 🙏🏽👊🏽

[@popsarocker]

If you're recording a show, please get a split before audio hits this processor.

[@devinlsheets_alphasound]

good advice, record lines should always be pre-processing, even pre-HPF.

[@audiotech3403]

MAGIC!!!!!

[@CarAudioInc]

It still sounds weird to me.

[@devinlsheets_alphasound]

We agree, having no feedback is indeed weird.

[@fernandokarmali1993]

@@devinlsheets_alphasound sorry guys, you’ve dropped this 👑

[@ionutmarion848]

👍🏻👍🏻

[@namalsounds]

❤

[@peteleoni9665]

I am a detective. You are not telling the whole truth.

[@mikets42]

Scam. AI can not exceed CR performance bounds. You do not need AI to perform dereverberation but Multi-Model Adaptive Filtering to track ever changing RIR. No, you can not use FFT. It seems that you use someone else's Open-Loop Delayless Subband Adaptive Filtering, without much understanding how it works and without crediting the algorithm's author. What a shame! The whole thing is so utterly disgusting.

[@devinlsheets_alphasound]

dang you caught us. You should go make your own rip-off version and compete with us.

[@IanHind]

Waves Feedback Hunter and X-FDBK.

[@jasonzdora]

Sorry but 500 bucks is way too much for a VST.

[@devinlsheets_alphasound]

Normally, yes. But this isn’t exactly just another compressor or modeled tube amp or whatever. There’s nothing else out there like it right now. Also, $500 isn’t that much to pay to get rid of feedback… if you’re a church for example, and you use it for five years straight, that’s $2 a week.

[@jasonzdora]

​@@devinlsheets_alphasound I agree with your sentiment, overall. Churches will definitely pay out the nose for "professional industry standard" AV equipment because they have the money from their collections. From a DIY sound guy perspective, its too expensive. They could sell over 10 times more if they reduced the price to 100. Much more realistic number and it would double their sales revenue.

[@devinlsheets_alphasound]

@@jasonzdora yeah maybe, who knows. I’ve never known a pro engineer to look at any piece of equipment that does something this mission critical and say that $500 is too expensive, and it’s already selling really well all over the world. It definitely is not priced for the DIY market but maybe someday, we’ll see.

[@1toonhead]

So it just the job of a gate and a equilizer. Maybe if operators learn how to mix properly and artist learn how to use a pa properly we won't be needing it. I will NEVER trust any ai.

[@_Thumbnail_]

Oh well, but how many events are there with (!) propperly skilled operators AND artitsts/narrators/moderators? These naturally won't need this plugin. But think of ALL the other events (from church ceremonies to vendor presentations) which often have bad acoustics, badly calibrated PAs and no skilled technicians at hand. THESE will profit enormously! And these won't even care about a certain level of artefacts (or even notice).

[@nikm3r]

@@_Thumbnail_ Even if you have the best people and systems and room, just having the peace of mind that you'll get no feedback is priceless.