MAnalyzer settings to optimize for short clip average accuracy
-
MeldaProduction MeldaProduction https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=176122 - KVRAF
- 14339 posts since 15 Mar, 2008 from Czech republic
I assume you are working with musical signals, right? Then just keep the defaults and use the averaging to get some average of specified time frame (like the 1-4 bars you mentioned). If you need higher resolution, use super-resolution, or increase FFT size (and probably also overlapping, so that the display isn't too slow). It's usually fine to keep everything default imho.
-
- KVRist
- Topic Starter
- 452 posts since 21 Jul, 2018
Specifically, though, the log and analytic modes... are they more accurate representations of our hearing or not? The manual says rather cryptically, for instance, that the analytic mode is more accurate, then says it's usually best NOT to use it for music with no explanation.
-
MeldaProduction MeldaProduction https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=176122
- KVRAF
- 14339 posts since 15 Mar, 2008 from Czech republic
No spectrum analyzer is a good representation of our hearing!! Brain does hell of a lot of magic in there! As for the log averaging etc., I can only repeat - it depends on the signal. Human hearing doesn't average nor smoothen, so there is nothing related to it really.
-
- KVRAF
- 3511 posts since 27 Dec, 2002 from North East England
Just popping back in to reiterate Vojtech's point above. Never use a spectral analyser to determine what you can hear. A good analogy would be somehow sonifying and listening to the Mona Lisa to determine what you can see when you look at it. No such sonification/visualisation process could ever give you a fuller overview of what you've just seen/heard, it could merely hint at what you already know with the very broadest strokes. Sometimes it will straight-up pollute the information you've gleaned from the appropriate sensory organ. E.g., "the analyser shows a peaky resonance at this frequency - I couldn't hear it before, but I can definitely hear it now I've seen it" - if you couldn't hear it before you saw it, chances are it wasn't a problem that needed fixing. Such tools are best used to detect problems you can't ever hear - excessive ultra-low/high frequency build-up, for instance.
Wojtech's point about averaging is interesting too. At ultra-high frequency resolutions/low time resolutions, the graph can be thought of showing the average of every frequency's occurrence over the size of the FFT window. Say, about 3/4 of a second with an FFT size of 32768 in a 44.1 kHz (i.e. 44100 samples per second) DAW project. Your ears of course heard everything happening in the last 3/4 of a second unfolding from moment-to-moment, not all bundled up into one instantaneous moment of time.
Imagine this sound is the exact same length as your FFT size. It's exactly 1 second long, so if your DAW is at 48kHz, that's an FFT size of 48000. https://soundcloud.com/chqtestsubjects/ ... on-example
While we hear the above, the FFT has no internal sense of time, so what it's 'hearing' and showing you the graph of is actually this. The average level of every frequency that was in the entire sound: https://soundcloud.com/chqtestsubjects/ ... -example-2
If you run both sounds looped through the analyser with FFT size set to 48000, they will appear exactly the same on the graph.
Wojtech's point about averaging is interesting too. At ultra-high frequency resolutions/low time resolutions, the graph can be thought of showing the average of every frequency's occurrence over the size of the FFT window. Say, about 3/4 of a second with an FFT size of 32768 in a 44.1 kHz (i.e. 44100 samples per second) DAW project. Your ears of course heard everything happening in the last 3/4 of a second unfolding from moment-to-moment, not all bundled up into one instantaneous moment of time.
Imagine this sound is the exact same length as your FFT size. It's exactly 1 second long, so if your DAW is at 48kHz, that's an FFT size of 48000. https://soundcloud.com/chqtestsubjects/ ... on-example
While we hear the above, the FFT has no internal sense of time, so what it's 'hearing' and showing you the graph of is actually this. The average level of every frequency that was in the entire sound: https://soundcloud.com/chqtestsubjects/ ... -example-2
If you run both sounds looped through the analyser with FFT size set to 48000, they will appear exactly the same on the graph.
-
- KVRist
- Topic Starter
- 452 posts since 21 Jul, 2018
I get all that in the abstract, but for me, it's a simple practical matter that any tool should always be optimized for the task at hand. If that weren't the case, it wouldn't have all the settings to begin with.
I had people tell me all the usual stuff about using my ears and not relying on visual tools when I was trying to get a waveform display optimized. Can't say how glad I am I kept pressing forward. That it is far more accurate than my ears for many common tasks is probably an understatement.
I fully accept that the tougher question of relative value of short/long sounds within an average may be much more complex, but I do not believe that questions like whether analytic smoothing or log mode better approximate human perception are unknowable. UnKNOWN, perhaps, but not unknowable, and that implies that one answer is preferable to another even if we have trouble deciphering which.
As to your example, I understand the futility of asking the tool to do something impossible, but for doing what it CAN do, I see no reason not to seek the highest accuracy for the given task. Some engineers swear that no visual tools should be used at all. I couldn't possibly disagree more. If I didn't already love MAnalyzer (specifically FOR it's flexibility), I'd be bugging someone else in THEIR subforum about how to dial theirs in. I'm here because it's a good, solid, useful tool, and highly customizable. That brings with it a certain degree of curiosity.
I should also note that there is a strong tendency among audio engineers to believe that a great many things SHOULD be perceived solely as sound. Not only is this notion unusual among the population at large which tends to be overwhelmingly visual, but of all the 4 major representation systems, auditory processing ranks dead last on my personal hierarchy. If anything, I always find it much more productive to blend modalities, and spread information among the senses.
A number of the most interesting breakthroughs in human perception have come down to this concept, though I'm sure each of those individuals just seemed crazy until they started proving they could "see" and "hear" things others could not as a result. I feel that way right now about my waveform display. It's downright magic in it's accuracy. People keep telling me I can't possibly notice the things it makes clear. In isolation, perhaps I can't, but in combination, having access to that degree of accuracy is a game changer.
I had people tell me all the usual stuff about using my ears and not relying on visual tools when I was trying to get a waveform display optimized. Can't say how glad I am I kept pressing forward. That it is far more accurate than my ears for many common tasks is probably an understatement.
I fully accept that the tougher question of relative value of short/long sounds within an average may be much more complex, but I do not believe that questions like whether analytic smoothing or log mode better approximate human perception are unknowable. UnKNOWN, perhaps, but not unknowable, and that implies that one answer is preferable to another even if we have trouble deciphering which.
As to your example, I understand the futility of asking the tool to do something impossible, but for doing what it CAN do, I see no reason not to seek the highest accuracy for the given task. Some engineers swear that no visual tools should be used at all. I couldn't possibly disagree more. If I didn't already love MAnalyzer (specifically FOR it's flexibility), I'd be bugging someone else in THEIR subforum about how to dial theirs in. I'm here because it's a good, solid, useful tool, and highly customizable. That brings with it a certain degree of curiosity.
I should also note that there is a strong tendency among audio engineers to believe that a great many things SHOULD be perceived solely as sound. Not only is this notion unusual among the population at large which tends to be overwhelmingly visual, but of all the 4 major representation systems, auditory processing ranks dead last on my personal hierarchy. If anything, I always find it much more productive to blend modalities, and spread information among the senses.
A number of the most interesting breakthroughs in human perception have come down to this concept, though I'm sure each of those individuals just seemed crazy until they started proving they could "see" and "hear" things others could not as a result. I feel that way right now about my waveform display. It's downright magic in it's accuracy. People keep telling me I can't possibly notice the things it makes clear. In isolation, perhaps I can't, but in combination, having access to that degree of accuracy is a game changer.
-
MeldaProduction MeldaProduction https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=176122
- KVRAF
- 14339 posts since 15 Mar, 2008 from Czech republic
Just a remark, as I didn't read everything - there is NO accuracy in spectral analysis, because humans just "hear" in a very different way. That's why I always suggest keeping everything default, just increase Resolution for vertical resolution, and FFT size x Overlap for horizontal resolution. But don't try analyzing scientific signals assuming you will analyze musical signals the same way, it would be a complete waste of your precious time!