Audio mysticism

[Zaphod (giancarlo)]

This musical knowledge becomes a problematic distraction in engineering. I remember reading an interview with Daniel Weiss (maker of very high quality digital processors), stating that he doesn't value listening tests during product design! He simply doesn't allow his ears to decide.

That might be ok considering the type of gear he designs, as most of this stuff can be measured objectively and doesn't need any perceptive or empirical approach. But can you imagine him doing a reverb this way? I can't.

Indeed. This anecdote just impressed me a lot, given the results he achieved. I grabbed this over at GS, where this quote definitely seemed to trouble several high end "fans". They definitely expected more crystals, meditation and golden ear superpowers being involved.

I respectfully disagree
It depends on what kind of dsp you are creating. Sometimes you need to understand what you are doing and what the recording engineer needs, it is not just a matter of magnitude /phase/ harmonic dist/ time constants, but a relationship between a lot of variables. Go figure if rupert was deaf.
Problem is if you need to create a surgical tool or a musical one. The reverberation example is good: you could create a perfect clone of a plate or a true space and maybe the math alone is helpful, or a new musical one and you need proper listening skills.
I think today you need to be "complete" in order to understand complex processes and to be competitive in a market where the team size is pretty small, unless you want to stick on a specific kind only because the business allows that
My listening skills helped a lot in the past, this is my specific case and experience. Today I could judge an eq just looking to a bode diagram, but just because I know what kind of phase/magnitude is required for a specific task, and in most of cases I'm still wrong (everything is good on the math side but the filter just doesn't sound as much good as expected)

[FabienTDR]

That was an extreme example, not a set of rules.

No matter how the designer comes to his conclusions, there's no such thing as magic powder. Every reproducible "effect" has a rather logical cause. Be it the bandwidth narrowing down with gain, EQ coeffs moving with gain, hysteresis effects, a reverb design cleverly avoiding artefacts, whatever. IMHO all this can be planned and predicted upfront (given adequate experience in the field of course).

You never surprised yourself imaging your own expectations while the effect really was bypassed? Or fooled by simple loudness changes? I certainly learned to not trust my ears too much, it hit me too often.

Reverb and dynamic convolution are edge cases. I'm talking about pure engineering tasks, not patch design (which rarely has a clearly defined goal anyway).

Do you debug your dynamic convolution engine by ear? (e.g. "hey there's a bug truncating impulse[554][134], let's fix it")


Even the analogue design process doesn't offer much room for magic. Only few designs really work, the closest to magic is spending time matching high tolerance components. Or maybe wildly trying out similar components having side effects the designer most probably doesn't fully understand (or can't measure). These effect can be nice, but the designer has no control over them anyway! It's take it or leave it. Try and error, no wizardry or anything particularly advanced. There's typically not much potential for magical variation, be it an algorithm or an analogue circuit.

[JCJR]

I never did much groundbreaking but was busy developing ideas for many years. The process was usually something like--

_ Get an idea of a function or process or end result

_ Imagine and scribble on paper building castles in the air how to accomplish the goal.

_ Build a prototype and usually find that the result isn't as good as was imagined.

_ Back to the drawing board to figure out what was wrong, or think up better ways to do the job.

_ Reiterative loop of test / refine / test until it is as good as its gonna get, or the time budget runs out and the idea must be at least temporarily shelved til later.

Was not uncommon to revisit either finished or unfinished work after months or years, and think up improvements which were not obvious the last time thru. If not bounded by time budget, many times ideas were declared finished because I was too confused to figure out anything else to try to make it better. Maybe "further inspiration after a rest" can be blamed on the unconscious continuing to work on the problem after the conscious mind gave up.

Actually maybe relied too much on unconscious problem solving. Debugging or design or data gathering or calculation "possibly not very productive busy work" while giving the unconscious time to solve the problem. Often solutions seemed to just "pop in the mind" rather than consiously solved step by step.

I'm fairly lazy about sustained effort on a problem. The most miserable tasks were those where a solution didn't just pop out on schedule, requiring sustained attention to too many details over days or weeks. Near painful to grunt out work step by step with no assistance from "flash inspiration" or whatever. Resentfully wondering, "Why did my subconscious go on vacation and abandon me to this mind-numbing monkey work?"

[stratum]

"Why did my subconscious go on vacation and abandon me to this mind-numbing monkey work?"

Probably you aren't having enough REM sleep. Might be because of too cold or too hot room temperature or sleep interuptions.

Just kidding - I sometimes similar problems as you have described and "the subconscious" never cares about them, it's not going to figure out its own misconceptions unless hitting to a wall, and then it can become pretty brilliant, after the fact In other times it works on automatic pilot without too much thinking and the results are OK. Is that unconcious problem solving? Well, probably it's just doing what you know already.

[nonnaci]

I find that sustained effort prims the intuition to solve / transcend existing problems when I'm not actively working on it (e.g. insight when jogging/eating/showering).

[Vertion]

[antto]

then you must also have the algorithm which can "zoom" into a single pixel and "enhance" the missing image information

i can think of an algorithm that can compress audio into about 260 bytes..
basically the algorithm uses a string with the audio name, e.g. "james brown - i feel good" and the algorithm uses LUTs to fetch the audio.. yeah it'll require extra space for LUTs covering every possible audio recording.. or fetching from the internetz

[Zaphod (giancarlo)]

That was an extreme example, not a set of rules.

No matter how the designer comes to his conclusions, there's no such thing as magic powder. Every reproducible "effect" has a rather logical cause. Be it the bandwidth narrowing down with gain, EQ coeffs moving with gain, hysteresis effects, a reverb design cleverly avoiding artefacts, whatever. IMHO all this can be planned and predicted upfront (given adequate experience in the field of course).

You never surprised yourself imaging your own expectations while the effect really was bypassed? Or fooled by simple loudness changes? I certainly learned to not trust my ears too much, it hit me too often.

Reverb and dynamic convolution are edge cases. I'm talking about pure engineering tasks, not patch design (which rarely has a clearly defined goal anyway).

Do you debug your dynamic convolution engine by ear? (e.g. "hey there's a bug truncating impulse[554][134], let's fix it")


Even the analogue design process doesn't offer much room for magic. Only few designs really work, the closest to magic is spending time matching high tolerance components. Or maybe wildly trying out similar components having side effects the designer most probably doesn't fully understand (or can't measure). These effect can be nice, but the designer has no control over them anyway! It's take it or leave it. Try and error, no wizardry or anything particularly advanced. There's typically not much potential for magical variation, be it an algorithm or an analogue circuit.

No because when I hear a problem, there is a problem. If it sounds good (but an analyzer tells me the opposite) I don't care as much as the other case.
This is a personal case. Yes I debug things by ear as first instance. When a problem arise I try to understand how to create an unit test, but the listening test is my final word on all things (engines, products, methods)

The process in such case is:
- tool/unit test for checking that apparently everything is fine (for example the bode diagram...)

->

- Listening test

All the times I was confident and I skipped the later we committed incredible mistakes

Said that, if I have to debug a new partitioned convolution algo, for example, my first unit test is based on automatic comparison with the expected result. But that's not enough, I never trust it completely, I have to "touch" with my hands for believing really. In a lot of cases something was wrong exactly on the expected result part!!!!

In my daily job the comparison action is still relevant, because there are many details very difficult to comprehend
I suggest this reading:
http://www.audiosignal.co.uk/Resources/ ... ent_A4.pdf

[Zaphod (giancarlo)]

(Yes it is old... But each age is raising new questions, we don't have an answer to everything)

[Vertion]

[nonnaci]

(Yes it is old... But each age is raising new questions, we don't have an answer to everything)

It was a good read nonetheless. From experience, psychoacoustic effects should not be ignored as elements of short-filter designs such as eq. are magnified when applied to long-filters. Case in point, pre-ring and pole resonances produce obvious colourations in reverb design.

[stratum]

I was a bit surprised to see that the article does not mention wavelets when considering simultaneous time/frequency analysis, since they predate the article roughly by a decade (1982 vs 1990). Therefore it would be natural to ask whether multiresolution analysis would reveal the audible effects mentioned quantitatively without needing subjective listening.

[nonnaci]

I was a bit surprised to see that the article does not mention wavelets when considering simultaneous time/frequency analysis, since they predate the article roughly by a decade (1982 vs 1990). Therefore it would be natural to ask whether multiresolution analysis would reveal the audible effects mentioned quantitatively without needing subjective listening.

This depends on whether human ears resolve acoustic events in frequency at the same temporal resolutions specified by the wavelet. And even so, it wouldn't adequately explain neural exhibitory-inhibitory based phenomenon such as temporal fusion of a direct and early sound-reflections, the effects of inter-ear coloration on sound-source localization, psychoacoustic masking.

[antto]

oh now it all makes perfect sense..
another developer who has genius ideas and talks about "God" and how his ideas were told to him by "God"
from previous experience, this will go wrong pretty quickly

[stratum]

deleted