| Columns Retired Columns & Blogs |
March 17, 2008 - 3:14pm
#1
Artifact Audibility Comparisons
| Loudspeakers Amplification Digital Sources | |
Analog Sources Accessories Featured | |
Music Columns Retired Columns | |
Show Reports | |
Features Latest News Community | |
Resources Subscriptions |
© 2024 StereophileAVTech Media Americas Inc., USA
All rights reserved
Thanks, Ethan. Fun and interesting files.
Nice 'cello snippet. Well done.
You are correct, the noise by itself is very annoying.
What I found most interesting is that once I learned how to listen to the noise (that is, what it sounds like, the periodic appearance) it was much easier to detect it. This is a great example of how experience and educating our ears allows us to perceive more.
While I think that you are absolutely correct that masking makes unmusical noise and distortion harder to hear specifically, I find that the absence of such noise can make a tremendous difference.
For example, the self noise of my Beyerdynamic M160 ribbon mics is pretty high. If I apply a lot of gain when recording this noise is present in the recording, although at a very low level. The only way to hear this noise on its own is to turn the gain of my Grace headphone amp almost as high as it will go. (Any of the music at this level would damage my hearing and probably the headphones).
Yet, if I use a good noise reduction plug-in to analyze and remove only this noise from the recording, the recording is now much more transparent sounding. I don't hear the noise as gone as I can't hear the noise at normal listening levels. Yet the music sounds vastly better when the noise is absent.
Thus I don't think the noise as audible noise is what is damaging the sound. I think the noise does damage to the rest of the sounds in a way that masks details - perhaps making everything fuzzy.
I think this is also why anything that makes the silences quieter (the blacks blacker) improves the sound of our systems.
As an aside, while I understand that you were illustrating the audibility of low level artifacts - is there any consensus as to what jitter actually does to the sound of a waveform? Does it simply add noise? Does it change relative timings of frequencies? What does it do?
Thanks again for the post and article.
Thanks. If you follow the link you can download a 192 kbps MP3 file of the entire concerto. And for free, not 99 cents.
I think this might be due to other factors. Not sure if you know how those NR programs work, but basically they're a 1,024 (or more) band noise gate. When you sample the noise, the volume level in each band is analyzed, and the gate thresholds are set right at the edge of the noise level. As you no doubt know, when used with more extreme settings to reduce noise by 20 or 30 dB or more, you hear a swishing sound not unlike low bit-rate MP3 files. At that point the gating is much more aggressive, and more actual music is being removed. Just as with MP3 files.
You know how I'm constantly talking about the audibility of comb filtering? Well, here's yet another instance where a series of many closely-spaced filters can affect the sound quality. With more normal amounts of noise reduction, and depending on the filter settings, this might make the music sound clearer and more open because some content is being removed. The NR program I use (Sonic Foundry's) has an option to hear only what is being removed. If your program has that feature too, try it and see if you hear music fragments. If so, that means some of the music is being removed along with the noise. And that could account for what you hear, versus a subliminally lower noise floor.
Yes, jitter adds noise artifacts. There are also timing errors, but as explained in my article the timing stuff is in the GHz range. So by definition that's not possible to hear. Versus the in-band noise which can easily be measured.
--Ethan
I will definitely download the entire concerto. Nice recording of the 'cello itself BTW, expressive and rich.
You are absolutely correct that if one gets too aggressive with noise reduction there can be artifacts. I find that it is better to carefully make multiple smaller passes rather than one big one. Listening to the residual (what is being removed) is indeed one good way to help tweak the settings. My opinion is that one should never hear the noise even be modulated by the music, much less parts of the music itself.
I have SoundForge 9 and Sony's latest noise reduction (I miss Sonic Foundry). It does a reasonably good job if one asks it to work hard. I find some other programs better however, such as Izotope's RX. It's noise reduction features are quite a bit better (and take huge computing resources).
I don't know how the various noise reductions work, but there appears to be more going on than a noise gate, even with oodles of bands.
For example, what do the all the settings do, such as mode and bias in the SoundForge version?
Additionally, good noise reduction works even when the music is too quite to mask the noise itself. Even better, the good ones can even handle noises louder than the music. Pretty wild stuff.
Any ideas how to test the effect of NR on musical signals?
Leaving aside the timing errors of jitter (if these are in the GHz I certainly can't hear them), what are the noise artifacts induced by jitter that we can hear? A DAC driven by USB sounds quite a bit different than driven by a coaxial S/PDIF, for example.
It's the same people and they're at the same shop in Madison, WI. The main difference is their paychecks now say Sony, and their tech support now costs us money.
Could be, I'm not enough of a math geek to truly understand the entire process. The first affordable NR program was DCart (they may still be around) and the company was founded by a couple of geniuses from Hewlett-Packard. I had a number of discussions with their head programmer guy and that's how he explained it, as a bunch of noise gates. It makes sense to me that this is the main algorithm. Just as a gate has a threshold and amount of volume reduction (most don't have to be only on or off), the same applies for all 1,024+ bands.
That's where all those bands come into play. With a normal gate, if you have a lot of tape hiss on a bass track you can hear the hiss come and go as the bass player starts and stops. Having multiple bands on wide-range music means that all bands not in use at the moment are muted, and when the other bands do get passed through they're masked by similar frequencies in the music.
Again, loud hiss over bassy music is easy to gate away with many bands. And with all those bands it's not just a bass below 500 Hz versus tape hiss above 2 KHz - the software can be very selective about what is gated. Even with frequencies very close to each other. 1,024 bands means each is only 20 Hz wide. Increase that to 16,384 bands and the software has very fine control over the process.
I admit I'm mainly theorizing, but it all makes sense to me. In my mind, that is.
FFT analysis in Sound Forge? Or FFT on the parts that are removed?
I can't imagine why cable format would matter. Related: I downloaded two different scientific papers on this today, and using measurements and listening tests both concluded that jitter is a non-issue for all the reasons I stated. The artifacts are just too darn soft for anyone to hear.
--Ethan
Interesting thoughts, Ethan.
I haven't been able to find much information on how noise reduction actually works. I found one reference that indicates that it splits the frequency bands as you describe and then applies expansion to the noise frequencies to drive them lower into the noise floor. This is more or less as you describe.
Fun stuff.
Thanks again for the samples and the discussion.