SLP 11.0.60 - Issue with Unmix Multiple Voices

SpectraLayers
1

I am working on a repair job of a live recording of a comedy show where they borked the audio in many ways :roll_eyes:

It’s mainly a woman and a man on the stage doing their dialog routine and several musical performances in between. The male also does some voice impressions.

The female voice sounds ok, the male voice sounds too dull - but can be improved via an EQ match process.
Both voices are slightly clipped - that I can restore too.
Recording is from the mono sum of the mixer - no separations of the mics, and underlayed musical cues have been too loud compared to the live audio (I was told it sounded good live on site).
It seems the audio engineer did not care about the video crew’s needs (and the crew was undermanned)

In the second half of the recordings the video crew forgot to start the recording from the mixer output for the first 25 minutes. So only camera mic recordings do exist. When they finally pushed the rec button levels had been raised and more clipping got done. Yippie.

Anyway…

To make things easier for processing the male voice I was trying to use the Unmix Multiple Voices. Registered both voices. Processing the unmix and getting 3 layers (2 voices and 1 non-voice)

As with the other Unmix modules I was expecting this to be non-destructive when all layers are active. Unfortunately this is not the case. On several places Unmix did delete parts of the voices - like it was gating them out. Wouldn’t be a problem if those parts would sit on the non-voice layer - but that isn’t the case.

So far it seems Unmix Multiple Voices isn’t ready for prime time yet.

1 Like
2

I have to say that I have not used any unmix multiple voices processing successfully; manual unmixing is the only success I have had…it is an arduous task and you aren’t going to get clean stems; you can “re-balance”/ re-mix somewhat…

I unmixed my main contrib and a group of passers-by (5 or 6 other voices). All voices captured on location in a vineyard with a senn MKE-2/ ew100 radio lav and a Senn K shotgun. It took me about 2 weeks to get some separation and ability to rebalance. The components do not sound very good all by themselves; the resulting re-balance is impressive.

I’ve also been able to pull out telephone conversations when the main contrib is on the phone…person on the other end via main contrib phone speaker…also, very impressive… also completely manual

I mean, thank goodness for the Transfer tool!!!

Frankly I don’t use any of the selection tools which auto-select from the spectrograph; I mainly use brush, transfer and clone tools…sometimes rectangle/ ellipse selection and time selection tools…not often, tho

I find pasting from magic wand, freq/ harmonic all have anomalies at the leading and trailing edges. IME, These anomalies take far longer to remove than just re-doing manual selections. Further, these anomalies appear to add some events which aren’t there in the original audio, and that just confuses me with my (mostly) non-destructive editing processes.

Most of the modules just aren’t up to it yet, IME
just like you are saying
maybe soon, I hope
that said, I’m still using 11.0.40 build 395

The module I use on just about everything is unmix noisy speech, which also isn’t perfect, but my goodness, is it a massive, massive time saver.

I love SL; it blows my mind every day,
the future is bright

2 Likes
3

Umix Multiple Voices works very well, even when both voices talks over each other. The main issue is that there seems to be some destructive filtering out going on. Sounds the tool can’t identify as either of the voices should get into the extra non-voices layer and not get deleted. It feels like a parameter behind the curtain has been set too restrictive.

iZotope RX does have a De-Bleed module for that which can remove the bleed from one mic into another. That could help with preparing such files for further processing in SpectraLayers. Both tools (RX & SLP) do complement each other quite well. For example declipping of SLP is not even close to RX - I can get far superior results in RX.

Yes, I do the same but I prefer to get as much work as possible automatically done by the tools themselves and then just have to finesse the results.

With shorter clips it’s ok to go in all manually, but for a 2 hour recording - no way - no one is gonna pay me for that amount of work accordingly.

Yes, I am using it to get rid of squeaking floor noises, audience coughing, leather couch noises, talent touching its lav mic several times - all sorts of noises very hard to get rid of when blended into human voice.

1 Like
4

To clarify: I wanted to increase the volume of the passers by
which is what I achieved. I’m not sure I understand ““bleed” from one mic into another”…I was on location with an omni lav and a shotgun; the mics picked up the scene

I think of bleed in more of a tape track bleed than mic bleed

I can’t stretch to purchasing RX

5

Can you please send me a link for the program you have that works as you said? Because for me, and for years, Unmix Multiple Voices has always been completely broken. It can’t even differentiate a male and female voices talking at different times, in a crystal clear recording. For me it absolutely never ever worked not even a little.

1 Like
6

@henrique_staino Do you have a short audio sample you can upload which suffers from this issue?

1 Like
7

It could literally be any of the hundreds upon hundreds of files with two voices I’ve worked with. In any shape or form, or recording condition, or degrees of dissimilarity between the voices.

I guess I could find one here, but it is just the same if you take any one you have there. I’ve tried it on so much stuff that I’m positive that any file you have in your HD would yield the same results.

1 Like
8

I tend to agree, it’s not really working.

I’m not convinced by the current Unmix Multiple Voices module. In my opinion, it actually did a better job in an earlier version of the module where it automatically found the voices. The new way of identifying the voices first is not optimal and is a regression in terms of the results.

2 Likes
9

FYI, just went back and did a comparison between SL11 and SL10 with Unmix Multiple Voices with the same audio file. I hear a superior result in SL10 - it’s not perfect either but it’s a LOT better than SL11. In my opinion, SL11 has clearly suffered a serious regression in this regard.

1 Like
10

Let me just say
I was reviewing a mix comparison from before and after using SL to separate multiple speakers in the same mics;
Two completely different scenes recorded with the same gear;
the rebalance is astonishing, lose the noise floor and magic can be created

my director was suitably impressed…I had thought is was a slight difference…the waveforms tell a different story

11

I have now tested the exact same sample with SL10 and SL11 - both are giving me the same results.

But what I found now is that “Register Multiple Voices” does crash a lot - I mean really a lot.

My system specs are probably at the top 1% of what people here are using for a DAW:

  • AMD Ryzen 9 5950X (16+16 Cores)
  • Gigabyte B550 Vision D-P
  • Corsair Vengeance LPX 128GB (4x32GB) DDR4 3600MHz C18
  • PNY NVIDIA RTX A6000 (Ampere) with 48GByte with Studio drivers
  • Seasonic PRIME TX-1000 80PLUS Titanium 1000 Watt
  • WD_BLACK SN850 2TB NVMe; PCIe Gen4 for system
  • Crucial P5 CT2000P5SSD8 2 TB for fast cache
  • 4x Crucial MX500 4 TB as 16 TB RAID 0 for big cache
  • LSI 8 port RAID Controller 8x 16 TB Toshiba Enterprise [MG08ACA16TE] as 104 TB RAID 5 for work files

Additional infos:
SLP 11.0.60
NVIDIA WHQL Studio driver 553.50

Audio file:
Mono 48kHz 32bit WAV
Length 2m 15s
2 Voices (male & female)

Logfiles after the crash do not show any hints - last entry is always
Project Samplerate: 48000Hz Device Samplerate: 48000Hz

Looks like the crash does prevent any entries added to the log.
Crash dumps got written though.

2 Likes
12

Yes. I’m sure results are highly dependent upon the source audio file too.

1 Like