Topic: Fixing a once off talker. Ideas?

[shadowfax1007]

I've got an old recording of mine which overall I love - except for one track which almost ruins the listening experience for me.
At the 3:14 mark you here a drunk girl talking nearby to me, picked up strongly on the right channel.

I've experimented several times over the years with various audio tools, but I've never really been able to reduce it in a way that doesn't harm the experience too much. I'm kind of hoping some more seasoned editing veterans might be able to have a quick look at this one and make some suggestions for me? Please.
Alternatively I've got RX10 if people have settings/tools to suggest trying.

I've posted a link to the MP3 of the track itself, but I have the whole gig in it's original WAV format if the edit is required overall.


https://we.tl/t-It3luggwy4

Thanks!

[Gordon]

Look at it in RX and figure out what her voice looks like (easy to spot).  Use the brush tool or lasso then spectral repair attenuate.  This was a quick attempt to tone it down.  It could be done better but you get the idea....


https://we.tl/t-SizfXMKKCb

[breakonthru]

Its REALLY hard to get conversation out of a recording as it’s not sharp and sudden and it’s in the same frequency range as the music

When it’s hard on one channel the first thing I do is subtly swap channels via long fades. The image will collapse into mono and be noticeable on headphones but sometimes not noticeable at all when played back on speakers

Perhaps some combination of the above, where at first you try to reduce it on the one channel as much as possible, and then swap the channels as little as possible to try to maintain a little bit of image, and use the slight channel swap to partially cover up the manual work you did to remove the tower, which is not without consequences to the rest of the recording

[nassau73]

Look at it in RX and figure out what her voice looks like (easy to spot).  Use the brush tool or lasso then spectral repair attenuate.  This was a quick attempt to tone it down.  It could be done better but you get the idea....

True. Work in the RX Spectrogram View. The voice of the "talker" will be viewable as concentric lines in the area where they are talking.

This is a process I follow when I'm willing to be truly anal about getting rid of the talking. It's very time consuming so judge what you are willing to commit. I've been going back to some of my old 24 bit files and "fixing" ones that I want to listen to without the talking and screaming. This is also the process I follow when the talking is interfering with the music or vocals. If I want to bring down crowd noise between songs - that's different.

Remember that the loudest part of the talker will be closer to the bottom of your view.

That in mind, once I locate the the concentric lines of the talker, I go to the bottom of the display and SELECT a small area using the BOX SELECT TOOL that's the rightmost tool in the icons.

Then I use the magnifying glass icon that has the dotted lines around it. This magnifies everything so that now it's easy to start highlighting individual lines of the talker. I use the paintbrush tool set to a size that allows me to select the lines of the talker without affecting the surrounding areas. I start at the bottom and many times have to select every other line rather than the lines next to each other.

Once those are highlighted, I use the harmonic tool to select the harmonics out of range of the smaller selection that I'm viewing. Now attenuate or delete in spectral repair.Then I do the rinse and repeat thing - go back to what I missed the first time and do it again.

After I did this for awhile, I learned that I could see the talker lines in the middle of a vocal or interfering with the music so it's possible to really clean things up.

On recordings that I want to listen to and have egregious talker intrusions and am willing to spend the time - this has been working for me. My computer doesn't have the resources to split stems, etc.

Good luck if you try this.

[robgronotte]

If the talk isn't over any singing, Ultimate Vocal Remover 5 is amazing, you can download for free online. I've used it so much in the last year that I consider myself an expert.

I would be happy to work on it. Would be great to let people hear the before and after.

[breakonthru]

It would be super valuable if you documented this process while doing it so we can all learn

[shadowfax1007]

Look at it in RX and figure out what her voice looks like (easy to spot).  Use the brush tool or lasso then spectral repair attenuate.  This was a quick attempt to tone it down.  It could be done better but you get the idea....


https://we.tl/t-SizfXMKKCb

This actually sounds much better. Still not perfect, but helps remove the harshness of it. I'll definitely be giving this a decent crack this week. Thank you.

[shadowfax1007]

If the talk isn't over any singing, Ultimate Vocal Remover 5 is amazing, you can download for free online. I've used it so much in the last year that I consider myself an expert.

I would be happy to work on it. Would be great to let people hear the before and after.

I'll do a general write up soon on my experience - but holy crap UVR5 worked great. It certainly wasn't a single click solution but after a few hours of tweaking with UVR5, Audacity and Izotope, I got a great outcome. Managed to speak the lead vocals and the "backing vocals" and blend the lead only back with the instrumental. Then pasted the exact clip back into my overall recording. You almost can't tell it was there to begin with. The intrusive vocal is totally gone now. The remaining actual vocals are ever so slightly muffled for like 1.8 seconds, but listening to the entire recording you can't really pick it. It's just like the singer has moved slightly from the microphone for a minute.

Fantastic bit of software!

[shadowfax1007]

So far anyone wanting the process I followed:

1. - Audacity
- Set 'Start' and 'End' markers around the problem section.
- Copy the problem section to a new track
- Export the new track only, ensuring to save in the same format and bitrate
- Save the project file with markers, or just keep the window open.

2. - UVR5
- Download the following models if not already installed: KIM1, UVR-BNE-4B, UVR-Deecho-Deverb
- Set process method to MDX-Net
- Set model to Kim Vocal 1
- Go to Settings > Advanced > Advanced MDX > Vocal Splitter
- Select model UVR-BNE-4B
- Select "Save Split Vocal Instruments" and "Enable Vocal Split Mode"
- Select "Deverb Vocals" and enable "All Types"
- Go back to the main screen, process your input audio.
- UVR5 will output individual stems, and various combined steams such as "Instrumental + Backing Vocals" or "Instrumental + Main Vocals"
- Listen to your output and see how it goes.
- You can experiment further with Segment Size and Overlap but the higher you go, the longer it will take and the more grunt your PC will need to have.
- Some people recommend using Kim Vocal 2 model if Vocal 1 model doesn't give the required result. In my case, I found no noticeable difference between the two.
- Make sure you export your audio in the same format and bitrate.

3. - Import back into Audacity
- Go back to Audacity
- Select the range set by your markers, it should come up in yellow and lock your selection to the exact time.
- Copy and paste your new audio over your main track
- Your new audio should be pasted over the old offending segment and time matched. Playback to confirm, and adjust if needed.
- Export your audio to a new master .WAV file
- Import the new master into RX Izotope or whatever editing software you use.

Enjoy!

[robgronotte]

The best model I've found in UVR5 is MDX-Net MDX23C-InstVoc HQ.

but I usually run the ensemble mode which does three models and then combines the results, and I pick the best of the 4.

the other two I use are VR Arc1_HP-UVR and Demucs v4 | htdemucs

[shadowfax1007]

The best model I've found in UVR5 is MDX-Net MDX23C-InstVoc HQ.

but I usually run the ensemble mode which does three models and then combines the results, and I pick the best of the 4.

the other two I use are VR Arc1_HP-UVR and Demucs v4 | htdemucs

If you're still willing to have a go, I'd be keen to see what you can do with this sample?
I've included the original .wav file slice and my version I ended up eventually.

[robgronotte]

Your old WeT link is dead and I don't see a new one.

[ol' dirty taper]

Would UVR work for someone who was eating chips in a bag underneath one of my mics?

[shadowfax1007]

Your old WeT link is dead and I don't see a new one.

Apologies - not sure what happened with my last comment.
https://we.tl/t-FTlMnRHAsW

Try that

[robgronotte]

Would UVR work for someone who was eating chips in a bag underneath one of my mics?

Probably not, it's somehow tuned to voices, not miscellaneous noises like that. Probably could be done on izotope RX but would be tedious if it lasted very long.

[robgronotte]

Your old WeT link is dead and I don't see a new one.

Apologies - not sure what happened with my last comment.
https://we.tl/t-FTlMnRHAsW

Try that

I checked them out.  Your fixed sample is longer than the original sample but I assume the problem is somewhere in the fixed part?  If so, that's really good, I can't even tell where the original problem spot was.  I wouldn't be able to do a better job than you did.

[shadowfax1007]

Your old WeT link is dead and I don't see a new one.

Apologies - not sure what happened with my last comment.
https://we.tl/t-FTlMnRHAsW

Try that

Cheers! Yeah the fixed sample is a bit longer, I did it as I was doing some blind testing of my edits. Wanted to see if I could pick it out or not based on different versions.

I checked them out.  Your fixed sample is longer than the original sample but I assume the problem is somewhere in the fixed part?  If so, that's really good, I can't even tell where the original problem spot was.  I wouldn't be able to do a better job than you did.

[goodcooker]

I have a recent recording I'm going to try this on. It's an instrumental ensemble - Hammond organ, guitar, bass, drums and sax - that the crowd in the tiny room just yakked really loud the whole time.

I hope to isolate the talking using the separate vocal stems method and discard as much of the conversation as I can.

Thanks for your write up about how you accomplished your edits.

EDIT - I took a short clip of the show from right at the beginning where the talking is especially bad and ran it through UVR5. Here's a clip of the results  https://soundcloud.com/roger-cox-7/robertwalter2024-05-03uvrcomp?si=822ae8d008194275ae17808e98ee4bcb&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing.

0-10 sec = raw track
10-20 sec = just removed noise (vocal track)
20-30 sec = just resulting music (instrumental track)
30-40 sec = raw track again to compare

It's better but it's not a silver bullet for this kind of thing. It definitely removed the worst of the talking and didn't degrade the music all that much. If you listen closely to the removed voices only you can hear a little of the organ that got removed too.

[robgronotte]

I have a recent recording I'm going to try this on. It's an instrumental ensemble - Hammond organ, guitar, bass, drums and sax - that the crowd in the tiny room just yakked really loud the whole time.

I hope to isolate the talking using the separate vocal stems method and discard as much of the conversation as I can.

Thanks for your write up about how you accomplished your edits.

EDIT - I took a short clip of the show from right at the beginning where the talking is especially bad and ran it through UVR5. Here's a clip of the results  https://soundcloud.com/roger-cox-7/robertwalter2024-05-03uvrcomp?si=822ae8d008194275ae17808e98ee4bcb&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing.

0-10 sec = raw track
10-20 sec = just removed noise (vocal track)
20-30 sec = just resulting music (instrumental track)
30-40 sec = raw track again to compare

It's better but it's not a silver bullet for this kind of thing. It definitely removed the worst of the talking and didn't degrade the music all that much. If you listen closely to the removed voices only you can hear a little of the organ that got removed too.

Which module did you use? The best one I've found is amazing but would be really slow to use on a whole show. Maybe 15x real time.

[EmRR]

I had pretty amazing results with UVR5 on a stealth DPA 4060 AB omni recording, old Buzzcocks show with many various people yelling in my ear. 
It did a better separation between music and vocal without leaving holes in the music as RX10 did in places.  (I haven't tried the same split in RX11)
Then I split the vocal with MDX-Net Karaoke 2. 

The vocal splits made fairly random tracks with vocal parts jumping from one track to the other, and some things that it turned into semi garbled separations that sound normal together.
In some cases, perfectly clean separation of a close loud talker. 
In others, judicious muting of various sections of the parallel parts got the talker out, or mostly out.  Some edits that used both parts of the split needed one part gain boosted to match, if it was a harmonically partial split section. 
One bonus was that I rarely had to mute both in the same spot, so the little bit of bleed ambience from other things doesn't go away as unnaturally as muting a single vocal would do.  Artifact of the process, but useful. 

It took the number of things I'd be trying to draw out in RX from something like 40 to 6, and of those some were best edited on one of the splits, others on both, others on the unsplit vocal.  Cut my workload considerably and allowed some edits I couldn't do with RX alone.

in the model downloads you'll find "MDX-Net Karaoke 2  Crowd HQ 1" which then shows options for "Crowd Only" and "No Crowd only"   Promising!    It did indeed work, though there are occasional pieces of the vocal that are also gone if you hit mute.  You can definitely drop it -6dB for some real improvement. 


What seems lacking is any set of descriptions for the numerous models - at least I can't find it if it exists. 

[rastasean]

Is this what ya'll are referring to?
https://github.com/Anjok07/ultimatevocalremovergui

What other plug-ins do you recommend?

[robgronotte]

Yes, that is what I use. Don't know about any plugins that are similar. Izotope RX10 has a similar feature but it doesn't work nearly as well.

[rastasean]

Thanks, robgronotte!

I played around with it using some music from youtube to grab just vocals and just instruments. It works pretty darn well!