SpectraLayers Demos/Tutorial (copied from another topic)

SpectraLayers
1

A fair enough request, I suppose, to move this to a new topic. Sorry for the duplication. My shortcoming that I am not aware of the “multiple topics” attempting to address this rather than simply complain. Any takers on improving the situation?

OP follows:

This may be a good area for a group of volunteers to jump in!

I have also found lacking the documentation of the primary use cases for many of the modules. I suppose that some of this may be that those who actually use it are busy doing the work rather than donating time to tutorials.

Personally, I feel I would benefit from one or two actual use cases for how each module might be used in real-world projects. Additionally, some comparisons of how one module might excel at a task that seems more obvious for another module.

This begs the question of where a compilation of scenarios and their solutions would originate, as well as who has the knowledge and time to produce these tutorials.

Seems a good start might be:

  • A collection of scenarios that users would like to (or have) tackled and demo of how the modules were used to solve the problem.
  • For each existing module, provide a more detailed description of what scenarios that module was intended to address and how to accomplish the completion.
  • Comparison between differing modules and how they might be used to tackle a task better than another module.
  • Guidelines in how modules affected by FFT size and amplitude settings affect the end result.

It would probably be legit to say that those who routinely use spectral editing already have intuition about this and insufficient time or will to do demos.

Equally though, many of us that use SpectraLayers may benefit immensely by being introduced to demos of how to solve problems we have not considered.

I would like to make better use of the Unmix tools, as well as improvements. However, I must admit some limitations in implementation. For example, after completing an Unmix, I often get results where I can clearly HEAR guitar and keyboards ending up in the same stem. What I’m missing is the HOW of software reliably identifying the two and separating them.

So, rather than pile on more work for developers (and increased cost for the product), how might a group of volunteers identify and implement demos/tutorials. This would ideally not involve copyrighted material and begs the question of authorized use for demo purposes.

Ideas anyone?

Any good pointers to existing demos/tutorials than we may be unaware of?

2

OK, so, how deeply have you delved into this forum? I’ve only been on this forum since end of SLP10 and logged my teething and shared plenty of my ignorance (the term: Transforms, immediately comes to mind) and findings as my time in SLP ticked up.

First step is: how well do you feel you can read the spectrogram?

I give an example: today I re-opened an early job where I spent at least one month working for 4-8 hrs daily on a 5 min location recording where the main contributor meets up with and has a discussion with at least six other people which was captured by the main contributor’s lavalier and boom mic.

I found plenty of human voice audio in the noise layers which could have been separated manually to the specific speaker’s layers;
AND
I found plenty of unwanted noise in the separated speech layers that “could be” sent back to the noise layer.

Nowadays I would not have missed those components that went to the “wrong” layers…partly because I went looking online and studied up human voice transforms. It was not glamorous research, yet has been invaluable. I didn’t do the work on the old job, because I already spent so much time on that one scene.

I won’t bang on about it more now, I have said plenty about that in the past on this forum and I also provided links. Further, I assume most SLP users interested in demonstrations of modules are not interested in manual work, they want to run modules and hope SLP delivers the results they imagine/ desire.

My point is, there is A LOT of great advice already in this forum, some of making sense of the Owners Manual or sharing real world solutions and workarounds…and plenty of bug reports.

I haven’t yet purchased SLP13, but I will soon; I’m just busy getting on with other things :slight_smile:

3

Frankly, I followed your request specifically to start a new topic about this. As you were quick to respond, I assume you are aware of this.

Obviously, the ability to “read” the spectrogram would be important, so I’ll assume that was meant constructively. I “feel” I can read it quite well.

My intent is not to dispute modules that don’t behave as expected or to discount problems. Rather, to address the vacuum of demos and tutorials, whether there is any interest in volunteers, what information may be useful, and in what forms. As they say, “A picture (or video demo) is worth a thousand words!”

The sad answer is that there may be insufficient interest with people of sufficient abilities and or time and energy to spend.

Your question goes directly to the point! How are people with questions about best use cases for SpectaLayers best able to find demos/tutorials/examples without hours of searching for use cases that don’t meet their needs.

It is exactly this need to “deep dive” that is one aspect of my point.

How might we all benefit from an example like Greg Ondo’s YouTube Cubase QA/demo sessions regarding SpectraLayers, for example?

4

Well, it was a suggestion :slight_smile: I need help too, I’m interested in process as well. At least we can have a back and forth of multiple posts on the same day with this topic in it’s own thread.

Sure, I have zero interest in arguments or condescension…I was simply stating I think that SL users benefit by being able to assess themselves as to their ability understand what they see in the spectrogram and relate said patterns to sound…clearly much is FFT sample size dependent, so there is that too.

I don’t know how others might approach the thankless task of researching…Personally, I put in the time, months on end adding up to a couple years now, and still learning daily. There are paid courses, yet I haven’t looked into them, so I only know about those courses from reading this forum.

Please don’t get me wrong, I’m just trying to be concise and often that type of communication can appear confrontational. I’m really just trying to help as best I can.

I’ve been searching internet for all kinds of advice over the years. Some hour long videos have left me with a single nugget that I eventually was able to really benefit from and helped me evolve my workflows.

I’m interested in your proposal, I do have a few ideas for videos for SLP, yet time to commit to that is still too tricky for me right now.

5

Damian Kearns has produced 12 parts so far about Spectralayers 13 - check out his site on Youtube:

@damiankearns6387damiankearns6387

Here is the first one:

SpectraLayers Pro 13: What’s New In SL13?

6

I appreciate the clarification, though am admittedly a bit confused. I thought I gave a bullet list of fair examples that might lead to something. I didn’t see where any of these were addressed.

I’m retired now, so time is not one of my major concerns, however wasting my time where there is inadequate interest or expertise, I don’t have much patience for. I might be able to contribute, however admittedly there’s a lot that might be better done elsewhere.

It seems that the first determination is whether there are adequate “informed” resources to contribute and whether there is adequate interest to be useful. I have plenty of “gripes” about a variety of things and make some attempt to make sure something useful comes from them.

It seems that step 1 would be figuring out how such an effort is not more of the same endless searching, especially for those less skilled in knowing what to search for.

In a way, I’m trying to address the endless complaints that I’m sure the developers grow weary of and exploring if there’s any way to address that without pushing for increased staff at Steinberg/Yamaha and subsequent increases in costs.

Could be a total waste of time! I guess we shall see.

7

Thanks! I believe this is a resource I’ve explored but I’ll follow up to be sure. Interestingly, if it’s what I recall, some of those same features have been reported as not working properly by other sources. I haven’t worked through them all so…

For some clarification, when I speak of tutorials, I’m somewhat referring to supplied audio project snippets, demonstration of how a module works, how to make the best of it, and the opportunity to use those snippets to repeat the exercise.

8

@Dave_B

Hi Dave,

Take a look at what Damian has. These are tutorials and all are new, using SL13. There are now 13 parts plus 3 more. He starts with the “What’s new” that I gave the link for, then Part 1 - Getting started with Spectralayers Pro 13.

Here is the description from his YT site:

This tutorial video series is aimed at helping people learn SpectraLayers, with a particular focus on audio post production. Most of the videos are meant to be watched on screens that are larger than a typical mobile phone, as they are part of a course.

Did you also watch the 4 videos about SL13 from Steinberg on YT?

This one was quite instructive, at least for me:

Workflow and User Interface Improvements | New Features in SpectraLayers 13

Like you, I need to see how the various parts of the program work, as the details in the manual are sparse.

John CB

9

Yes, I have verified that I watched most of these in their entirety and I do find them decent, though primary focused on new features in SL13.

More of these would certainly be quite welcome.

For the bullet points I listed above, it may be fair that I begin with some ideas of my own, as well as some of the shortcomings in how I have attempted to use some of the modules and the limitations I have encountered.

I mentioned one aspect of Unmix Song, for example, where guitar and keyboards (whether synth or piano) end up in the same layer or sometimes other. In such circumstance, I can clearly discern by ear the difference between the two and can both audibly and visually identify the two but have not been that successful in using other modules to assist in isolating the two into their own layers.

I’ve also found that trying to manually select the instruments is difficult for using “cut to layer” is problematic and it is extremely tedious work. Sufficiently so that isolating them for a remix is quite difficult and time consuming.

I’ve had very good results with separating vocals, crowd noise, bass guitar and drums, but had a lot of difficulty with instruments existing within the same frequency ranges. I have not had a lot of success with separating multiple vocals but have not worked much with the new modules for this.

I attempted to separate vocals from a business meeting from a room with very bad acoustics with mixed results, especially with multiple simultaneous speakers. I’ve tried various approaches with multiple modules without much success.

Granted, my use of the program may be stretching the boundaries of what is possible and may simply be unreasonable expectations.

Overall, I find spectral editing fascinating but could use much more detailed real-world use of most of the modules.

I’ve seen a lot of comments regarding lack of tutorials and demos and share many of these experiences, thus the ideas to explore how to better show typical use cases.

Thanks!

10

Assuming you mean dialogue here, rather than sung vocals. Which modules did you try?

With regard to tutorials, yes they can help and we could have more of them, but you might be better off posting short descriptions of specific examples of the processing you are trying to do with SL, including also uploading a short zipped extract of the problem audio. There’s a lot of expertise in this forum and others may well be able to point you in the right direction - but only if they can actually hear the audio you are referring to.

For example, try uploading short extracts of the guitar and keyboards which ended up in the same layer, and the spoken voice(s) from the business meeting.

11

Dave, I’ll provide a few brief tips which may help. I’ve been doing spectral editing professionally since 2004.

  1. You’ll benefit from learning about the underlying methods which comprise spectral editing. Todd spent some time doing this many months ago and it greatly assisted his results. This will help you when selecting or adjusting the FFT windowing parameters to achieve optimal editing. That’s what Todd’s talking about when he mentions the visual spectrogram.
    It really helps if you understand how spectral transforms actually work. Despite the innovative interface offered by SL, spectral editing is not like editing audio using normal methods. You are effectively editing audio which has been transformed into graphic data, then transformed back into audio data when the editing is applied. The factors which affect the accuracy of the display and the applied edits are not the same as the factors applicable to normal audio editing.
  2. Unmixing using AI models is not spectral editing. instead, it comprises 'black box" models trained on audio files. The models used in SL were trained by 3rd parties, then imported in SL and tweaked further by the SL developer. Unlike spectral editing, you dont have the option of editing parameters, beyond the “quality” settings, which really just offer a compromise between processing time and the quality of the unmixed result.
  3. It is extremely difficult to unmix instruments, vocal parts or sounds which are very similar. This is partly due to the nature of the source materials used for training and the blended context of the audio you want to unmix. For example, if two voices sound very similar, how is a machine model meant to even know two different voices are present in the audio? Unmixing models dont know what they are unmixing. There’s no thought involved, no awareness. This is also why models have trouble separating some guitar sounds from keyboard sounds, or horn sounds from keyboard sounds. Humans have the same problem when listening!
  4. Difficulty distinguishing between sounds is the main reason many “stem separation” tools only offer 4 stems - vocals, bass, drums and other.
  5. Due to the nature of SL as a desktop application, the developer cant do frequent updates of the including unmixing models. He can tweak the included models a bit to fix bugs, but adding new models has historically only happened when a new version is released (ie once per year). In this regard, it’s a good idea to be aware of alternative unmixing services available online, which are more frequently updated and often offer models which can separate sounds or parts which SL current models havent been trained to unmix.
12

Sorry, do you mean compromise? I’m supposing you meant compromise :slight_smile:

As far as separating instruments, orchestral or layers of same/ similar timbres; harmonic series has to be considered as it is the hinge-pin of “music” visualized on the spectrogram. I would not say my personal understanding of harmonic series is very deep. We’d suppose that it should be possible to design some formulaic processes to harmonic series because it is (mostly) constant. Understanding harmonic series and applying it to manual unmix workflow should help.

And then we have the layers of timbres, though, which set a very high bar for unmixing afaiaa. What if we have a whole bunch of instruments playing in unison? Can we expect unmix modules to separate those timbres? Triple-tracked guitars? ADT vocals? My goodness, seems like an easier path is just to replay the parts, no? Until we can train unmix modules reliably for certain timbres, this will remain a sticking point for unmixers.

These basics of sound and music might seem a digression, yet it is the basics which can aid users in a deeper understanding which can lead to solutions via workflow invention.

I have to say I do not understand “how spectral transforms actually work”. Personally, I use pattern recognition for visuals which correlate to sound and get busy selecting and sending selections to layers. Using my methods I do not yield quick results; most of my work in SLP is extremely time consuming. The results, though, I’m often amazed by.

ok, I best get back to EQ work zzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzz

13

yes, I meant compromise. (now fixed)
Harmonics, timbre - these are easy for humans to hear but hard for algorithms to separate unless they are trained on that basis.
Training stem separation models isnt easy or cheap. To do it really well, the source files need to comprise the unmixed audio and the mixed audio.
For example, the algorithm needs a mix containing the violin part, and it needs the violin part before it was mixed. But one example is insufficient. Many, many examples are needed - the more before/after examples, the better the resulting model. The server time for the training quickly becomes expensive, but the real problem is having good source material. The workaround of adding synthetic data is never as good as using original unmixed tracks with the mixed version. This is the real roadblock to having models which work with the huge variety of music and sound recordings that professionals must edit.

But yes, i completely agree that the more we understand our source material and tools, the better the results we can achieve. It is always time consuming to do excellent work when fixing audio anomalies and SL is no exception. The joy is the capability to achieve results which were not even remotely possible just a few decades ago (spectral editing) or just a few years ago (stem separation). The work requires patience, but is very rewarding and persistence does pay.

14

That has been my experience as well. In addition, I have to balance the value of learning the tools, the substantial time and effort required, etc. against the value of picking a challenging source material to work with in comparison the value of the results.

While I have not delved deeply into FFT algorithms implementation, I do have an appreciation for the difficulty, particularly in areas described in the last few posts.

I’m not meaning to discount the difficulty or the work that has been done. I’m left with the notion that if I can clearly see on the spectrogram and hear with my ears, their ought be SOME way to have to tool assist in separating them. It’s not lost on me that perhaps we are just not there yet technically, or the cost is simply not worth the effort.

I appreciate the offers to help if samples are posted. I can also appreciate, as I said previously, that many/most of the people with the appropriate skills to educate the rest of us, have actual work to do that earns revenue.

15

Oh man, I wish! Maybe there is a payday upcoming after of release of my documentary…somehow, some way. I certainly won’t get in equivalence to the time I put in! And ot that I should ever claim to attempt to educate anyone.

16

Just think of the exposure you will get!

Bet you never heard that one before! :slight_smile:

17

@Dave_B I’m not sure you saw my post above: SpectraLayers Demos/Tutorial (copied from another topic) - #10 by stingray

18

I did and appreciate the notion of posting actual audio for assistance. Didn’t mean to ignore you.

I’ve watched most of what I could find, and while helpful, have also found many not really addressing issues I have stumbled across in my learning efforts.

This is why I posted the bullet list originally.

Being retired, I have some time available to offer some help, but also recognize my limitations in some areas, as well as consistent comments about lacking information. Whether I want to put myself in the crosshairs of, “You didn’t address my issue(s)!” is another matter! :wink:

19

Exactly, that’s why it’s often best to post actual audio examples here. Nothing is more difficult to comprehend than words describing the finer points of sound and audio processing! Each case or set of cases is unique and may require unorthodox processing techniques. What works with one set of dialogue files may not work with another.

20

Got it! But I also have to consider that much of the material I’m using is not my material to post on a public forum, so there’s that to deal with as well.