Echo/delay cancellation possible as an unmix feature?

I had quite a few requests to remove echo/delay from a recording. Mostly it comes from wrong audio setups (e.g. in Zoom or Teams meetings). The delays can range from 50 up to 500 ms. I don´t know any software that can handle such problems and remove those echoes within the recorded sample. I heard a few examples from an audio engineer who seems to have a solution but of course he refuses to tell me how it´s done. It must be an own creation with some clever DSP. Wouldn´t such a feature be something for SpectraLayers? It would find into the context of all the unmixing possibilities. What do you think?

I have a lot to say here.

First of all, I agree(meaning the process could be improved), second of all, removing delays/echos can already be done. If it is 500 ms, then use the unmix levels process (this is assuming that the audio of the volume of the delay/echo gradually descends/declines in a linear fashion). If the delay is somewhere between 50-100ms then (believe-it-or-not) the select similar feature works very well for this (I’ve done this for a delay before and it surprisingly worked really well, especially when you lower the ratio).

I do agree that select similar could be improved to add rhythm/beat/tempo option along with a delay/echo option. However there are instances where (for example a zoom meeting where the FEEDBACK of the delay is so high in amplitude) you would have to manually select the outline of the voice and invert the selection and cut to new layer.

1 Like

I don´t believe that you can manually remove echoes (especially when the also have feedback) that occur when new words are spoken. And if the speaker is not slow you quickly get a mess with overlapping echoes. You can´t just select manually the outline of the voice because it does contain more than tonal elements. The select similar won´t detect those overlapping echoes with newly spoken words. The select similar is a bit of hit and miss for me. I tried it once on a simple drum loop to see if it find all snare hits. It does but refuses to find the last one. They were all identical. Unmix levels also does not work when echoes overlap. You would eliminate a lot of the rest you want to keep.
I think the solution would be an enhanced version of “find similar” combined with some DSP trickery like spectral inverting, adaptive filtering, convolution or something else (don´t know if AI could help in such cases).

Yeah, I agree.

If its possible, can send me a small snippet of what you are trying to achieve? It’s kind of hard to fully agree that it is not possible to remove echos/delays when I’ve done it numerous times myself (using various techniques).

I sent you a PM

2 ways to go about this.

You can manually select (using the time selection tool) each syllable and use the select similar (didn’t experiment much but sure enough it worked for me).


Use the reverb reduction process. The steps are to duplicate the layer then use reverb reduction then duplicate the unprocessed layer then phase invert and merge the layer then reverb reduction process again.

The first step is a little tedious but gives best results (especially if you want it to sound natural). The second step is a lazy way and gives mediocre results. However that audio example is definitely doable.

1 Like

Thanks Unmixing for your suggestions but I can’t get the clean result I sent you.
My initial post was more like a feature request for Robin to maybe implement such an unmix feature if possible in a future version.

@Amberfields okay, but I suggest that you post the same audio example you sent me here (in this topic) because the main developer might read this and dusmiss the topic as solved and assume there’s already a solution. At least if you post it here, it would give insight (to the main developer) into what can be improved.

@Robin_Lobel: Do you think removing delays/echoes from a vocal recording is possible as a new feature (please see my first post in this thread)? If you need some examples I could send them via PM (I don’t want to put them here).

That’s certainly something that could be explored for a future process yes. Adding it to the wishlist !

1 Like

If you’re looking for echo cancellation/demixing then probably the easiest accessible AI de-echo is currently either the one built into UVR5 or the online version at

1 Like

Thanks for the info about UVR5 and the online version. I have UVR5 but could you please tell me which model it is? I haven’t found a de-echo MDX-net or Demucs model in there (checked also the downloads). Thanks in advance.

It might be one of the models you have to download or a VIP model.

1 Like

I tried several recordings with these models and got quite a few good results. The models struggle a bit when the delay time is very short similar to a slapback delay or even a bit shorter. I got a really good result from a recording which has feedback delay (compared to 1 single echo) and a lot of reverb. Thanks again for your hint using UVR5. So AI is capable of doing echo cancellation and I really hope that it will find its way into SpectraLayers as a de-echo process.

1 Like

Essentially Reverb is just very short delay, so sometimes you need to de-verb and de-echo, but it all depends.