Intersample overs

I am starting a new thread regarding this topic, because don't want to pollute the original dac talk 2 thread.

Anyway, pasting from my post in the original thread:
----
Motivated by @skem 's profile post, I am moving one interesting topic to this thread.

Roughly two years ago, benchmark posted about intersample overs and argued their dacs addressing this issue in their blog.

https://benchmarkmedia.com/blogs/application_notes/intersample-overs-in-cd-recordings

I don't like any dacs from benchmark (dislike dac1 and dac2/3 for different reasons though), but their reasoning regarding the issue sounds valid to me. It did not contradict to the recent AES talk which @atomicbob introduced.

http://www.aes-media.org/sections/pnw/pnwrecaps/2018/oct2018/

For this reason, I've applied a little bit of digital attenuation with foobar for years. But can't tell very strongly whether or not gains exceeded losses (lost bit resolution). Do this mostly for the peace of mind. Haha.
----

It seems that everybody agrees that this issue exists.
What might be controversial is that whether or not this needs attention of end consumers -- both in terms of likelihood and seriousness. Also, it's reasonable to guess that benchmark's blog posing, oops "white paper", might implicitly contain lots of intentional or unintentional biases.

According to the AES talk (linked above), different dacs behave differently. So overall perceptional results may vary.

Then, why not testing by ourselves?

I chose three random tracks from three different albums among my collection, which all I love as great pieces of music. Release years of these three albums are 2014, 2015, and 2017. 2017 one is remastered.

Here are analysis results via MasVis which I frequently use to investigate tracks.
Also see their guide for graphs: http://www.lts.a.se/lts/manual

Track 1:


Track 2:


Track 3:



I won't say all these suffer from intersample overs -- like I said it might be related to dac chips and implementations. But all of them have dangerously hot parts. For example, the least seemingly hot "sample 3" hits the 0 dbs >150 times according to the histogram.

These tracks are uploaded to the google drive for those who want to test with own ears, gears, or matlabs lol.

https://drive.google.com/open?id=1eLZU8ARRyXbaXkiReS_e2E33rxd1VREt

Again, I have no clue how bad this phenomenon would be with the real tracks and real gears. May need odac to correctly hear them haha.

PS. I really chose tracks randomly from recent years. And all the three satisfied "hot" criteria. f**k loudness war.

 

Here’s a reddit on the topic that’s useful.

Again, my personal view is that these are real but exists for such short blips in the music that they aren’t worth worrying about (of order 1-10ms/song cumulative), especially if you didn’t notice them already. Using digital attenuation to solve the problem might be nice if you have a 20-effective-bit DAC with room to spare, but if badly done you might easily lose more than you gain.

 

6db is one bit, I was recently reminded, so realistically 99.9999% of redbook is fine too, because the only track I have that might possibly be dynamic enough to use all 16 is va pensiero from verdi's nabucco and I ain't too sure about that either.

Now, whether the modern int->fp64->int digital attenuators sound affect the sound or not, that's another question.