|www.ethanwiner.com - since 1997|
Artifact Audibility Comparisons
by Ethan Winer
Debate rages in both audiophile and professional audio circles about the importance of low-level artifacts such as distortion, jitter, quantization noise, and summing errors in DAW software. This article addresses the audibility of very soft artifacts, and includes Wave files you can download to discover for yourself at what volume level these artifacts can be heard. I also suggest watching the Artifact Audibility portion of my AES Audio Myths workshop video. This LINK goes directly to that portion of the video.
With modern digital devices jitter is typically 110 to 120 dB below the music, even for inexpensive consumer-grade gear. In my experience that is far too soft to be audible. Indeed, this is 20 dB or more below the noise floor of a CD! Even though jitter is a timing issue, it manifests as FM sidebands - noise-like artifacts that are added to the music. Depending on the nature of the jitter the sidebands may be harmonically related or unrelated to the music. The spectral content can also vary. Both of these affect how audible the jitter will be when music is playing because of the masking effect. Absolute volume also affects audibility - the louder something is, the more detail we can hear.
Masking is a well-known principle by which a loud sound can hide a softer sound if both sounds have similar frequencies. This means you can hear treble-heavy tape hiss more readily during a bass solo than during a drum solo. Cymbals and snare drums contain a lot of treble, so that tends to hide the hiss. Another factor is that our ears are most sensitive to frequencies in the treble range around 2 to 4 KHz, so distortions that lie mostly in that range will be more noticeable and more objectionable than artifacts at lower frequencies. Especially when the music does not contain much energy in the same frequency range. Likewise, some types of harmonic distortion are more audible than others - buzzy distortion containing mostly higher-order overtones will be more noticeable than distortion that favors lower harmonics. And even harmonics are generally less noticeable than odd harmonics because they are octave multiples of the fundamental frequency. Intermodulation distortion (IMD) typically contains both low and high frequencies, depending on the frequencies present in the music. Some IMD components are not related musically to the fundamental pitches, so IMD is usually more noticeable and dissonant sounding than harmonic distortion. But it still has to be loud enough to hear, otherwise the spectrum is irrelevant.
In fairness, some people believe there's more to the audibility of jitter than just the added artifacts. Similarly, quantization noise occurs when reducing 24-bit audio to 16 bits if dither is not applied, and some people believe this can affect things like fullness and stereo imaging. However, fullness is a change in frequency response that is easily verified. And good imaging, in my opinion, is related more to room acoustics and avoiding early reflections than low-level distortion or microscopic timing errors at radio frequencies. THIS article explains what I believe is the main reason people sometimes hear changes in fullness and imaging even when no change is likely or easily explained.
One obstacle to devising a meaningful test of the audibility of artifacts is creating them artificially in controlled amounts. Real jitter occurs at extremely high frequencies. For example, jitter of 1 nanosecond equates to a frequency of 1 GHz. Yes, GHz - that is not a typo. Likewise for distortion, which adds new frequencies not present in the original material. It's easy to generate a controlled amount of distortion with sine waves, but not so easy with music that contains many frequencies at constantly changing volume levels. Fortunately, distortion, like jitter, can also be expressed as artifacts some number of decibels below the music. For example, 0.1 percent Total Harmonic Distortion (THD) means the sum of all artifacts is 60 dB below the music. Every factor of 10 is equivalent to 20 dB, so 10% > 1% > 0.1% = -20 dB > -40 dB > -60 dB. When isolated from the music, artifacts 60 dB below the music might be loud enough to hear without raising the volume unnaturally. But when music is playing, the music masks the artifacts making them much harder to notice.
Since most people do not have the tools to prepare a proper test, I assembled a series of 16-bit 44.1 KHz Wave files to demonstrate the audibility of artifacts at different levels below the music. Rather than try to artificially create jitter and the many different types of distortion, I created a nasty sounding treble-heavy noise and added that at various levels equally to both the left and right channels. The spectrum of the noise is shown in Figure 1 at left. Since this noise has a lot of treble content where our ears are very sensitive, this biases the test in favor of those who believe very soft artifacts are audible. That is, the noise should be at least as obvious as distortion or jitter artifacts that occur naturally, if not more obvious. So if you play the example file having noise at -70 dB and can't hear the noise, it is unlikely that naturally occurring artifacts the same volume or softer will be audible to you.
To make the noise even more obvious, again favoring those who believe very soft artifacts matter, the noise pulses on and off rather than remain steady throughout the music. In all of the example files the noise pulse is about 3/4 of a second long, and restarts every two seconds. The first pulse starts two seconds into each file and lasts for 3/4 second. The next pulse starts four seconds in, and so forth. These test files are not presented as a challenge where you have to guess where the noise occurs. Rather, I'm telling you where it occurs and for how long. My purpose with these tests is to educate people, not trick them. Click the files below to play them, or right-click and use Save As to copy them to your hard drive for repeated playing (recommended).
The noise.wav file (70 KB) is the noise burst by itself, so you can hear it in isolation and know what to listen for when the music is playing. The level is at -20 dB rather than 0 because it sounds really irritating. I don't want you to lunge for the volume control when you play it at a normal volume level!
The concerto-40.wav file (4 MB) is a soft passage from my Cello Concerto with the noise mixed in at -40 dB. Since this passage is very soft, mostly around -25 and peaking at -15 dB, the noise is only 15 to 25 dB below the music. Everyone will easily hear where the noise starts and stops.
The files concerto-50.wav, concerto-60.wav, and concerto-70.wav (4 MB each) are the same as above but with the noise mixed in at -50, -60, and -70 dB respectively. In the -70 dB version the noise is 45 to 55 dB below the music. Note that there is a slight noise that occurs naturally in this piece at around 8 seconds. This noise is in the original recording and just happens to sound like my intentional noise! I think it was one of the violinists turning a page of the music.
The file men_at_work1-40.wav (3 MB) is a section from the "industrial" pop tune Men At Work I wrote for one of my company's educational videos, with the noise mixed in at -40. I had planned to create other versions with the noise at ever-softer levels as above, but it's barely audible (if at all) even at this relatively high level so I didn't bother.
The men_at_work2-40.wav file (4 MB) is a different section from the same pop tune that's slightly softer sounding, which potentially makes the noise at -40 dB a little easier to notice.
It's well known that we can hear music and speech (program material) in the presence of noise, even if the noise is louder. I've seen estimates that we can hear 10 dB below the noise floor, which in my experience seems about right. Of course, the spectral content of the noise and program affects how much the noise masks the program, as explained above. But then I saw a post in an audio forum where someone claimed he can hear artifacts 40 dB below the noise floor of analog tape while music plays. So to test this for myself - and for you - I created a series of files you can download and play through your own system. All of the files below are 16 bits, 44.1 KHz, 1.6 MB in size.
The tones+noise file contains pink noise and a pair of test tones mixed at equal levels. The test tones portion contains both 100 Hz and 3 KHz at the same time to be more obvious to hear when they turn on and off.
The tones-10.wav file is the same pink noise and tones, but with the tones 10 dB below the noise. It's still easy to hear where the tones start and stop in this file.
The tones-20.wav and tones-30.wav files are similar, but with the tones 20 and 30 dB below the noise respectively. I can just barely hear the tones in the -20 file, and I doubt anyone can hear them in the -30 version.
To include typical program material I also created files using speech and pop music. The file speech+noise.wav contains these two signals at equal levels. It's easy to understand what is said.
The speech-10.wav and speech-20.wav files are similar, but with the speech 10 and 20 dB below the noise respectively. When the speech is 10 dB below the noise you can hear that someone is talking, but you probably can't make out what's being said. When the speech is 20 dB lower than the noise, it's all but inaudible.
The last two files are even more typical, mixing the noise and test tones behind pop music at 40 dB down, similar to the hiss level on a cassette tape. The tones in this test also pulse on and off every second, to test whether artifacts such as distortion can really be heard if they're substantially softer than background hiss. In the music_tones-40.wav file the tones are at the same -40 level as the constant noise. In music_tones-50.wav the tones are 10 dB below the noise at -50. I can just barely make out where the tones start and stop in the -40 version, and when the tones are 10 dB below the noise at -50 I doubt most people will hear them. And this is with the noise only 40 dB softer than the music, versus the -90 dB noise level of 16-bit audio on CDs.
I'm confident that these test files bust the myth that anyone can hear artifacts 40 dB below a typical noise floor. However, it's important to understand that there are many types of noise. The audibility of artifacts and tones below the noise floor depends on the frequencies present in the noise versus the frequencies present in the program. Noise containing mostly high frequencies will not mask low frequency content, and vice versa.
My hope is that these tests will put into proper perspective exactly how noticeable artifacts really are at various levels below the music. Especially for people who have no way to readily test this for themselves.
Entire contents of this web site Copyright © 1997- by Ethan Winer. All rights reserved.