|www.ethanwiner.com - since 1997|
Does Dither Really Matter?
Conventional audio wisdom says that dither is required whenever you reduce the bit-depth of an audio file. Typically, when reducing a 24-bit mix to 16 bits for putting onto a CD. Dither is a very low-level noise that's added when reducing bit depth, and by definition it's about 90 dB below the music when reducing to 16 bits. Most people would have a hard time hearing noise that's 60 dB below the music, since the music masks the noise as demonstrated in my article Artifact Audibility. Yet if you ask a dozen audio recording engineers if dither is necessary when going from 24 to 16 bits, every one of them will say Yes. Even the manual for Sony Sound Forge claims dither is important:
If you want to burn a 24-bit audio file to an audio CD,
dithering will produce a cleaner signal than a simple
Some engineers even argue over which type of dither is best, claiming this algorithm is more airy sounding that that one, and so forth. But just because everyone believes this, does that make it true?
To be clear, using dither is never a bad thing, and it can sometimes help on soft material recorded at very low levels. So I'm not arguing against using dither! But I've never heard dither make any difference when applied to typical pop music recorded at sensible levels. I'm merely pointing out that not using dither is never the reason a newbie's mixes sound bad.
To test this for myself - and for you - I created THIS set of test files (12 MB Zip) containing both truncated and dithered versions of the same sections of my pop tune Lullaby. These are the exact steps I followed:
I started with the Lullaby example from my AES Audio Myths Workshop video on YouTube. I rendered it from SONAR at 24 bits and extracted four short sections. I then alternately dithered and truncated each section down to 16 bits, and renamed the files to hide their identity. So lullaby a/b are the same part of the tune, with one dithered and the other truncated. Same for the file pairs c/d, e/f, and g/h. Your mission is to identify which file in each pair is dithered and which is truncated, and email me your choices (email link is on my home page). The dithering was done in Sound Forge using High-pass Triangular dither with High-pass Contour noise shaping.
Added March 3, 2016: I replaced all eight files in the Zip above with new versions.
Added January 25, 2010: The files linked above and the explanation are new, replacing the original file excerpts that switched randomly between truncated and dithered. Some people fairly criticized that as an invalid test, because you're not able to hear the same music with and without dither. So with the new files this is now a valid test.
Added January 19, 2008: A friend and I did some extensive listening tests to compare dither, jitter, and A/D converter quality. I posted my report in Lynn Fuston's 3dB audio forum. But that forum is now gone so it's repeated here:
Posted January 19, 2008
Out forum friend Grekim Jennings came to my studio yesterday and we spent a fun four hours playing with our toys. Grekim is a fine musician, and you'll hear his guitar playing in the example files linked below. We did three tests in all:
- Comparing dithered versus truncated 16-bit files at loud and soft playback levels
- Listening for jitter by comparing my Delta 66 sound card with its own clock versus clocked by Grekim's Apogee 8000
- Listening to instruments recorded simultaneously through a SoundBlaster X-Fi sound card and his Apogee
For the dither test I loaded into Sound Forge a 24-bit mix Grekim has been working on. I created two 16 bit files from that, one dithered and one truncated. I then played the two 16-bit versions alternately while Grekim listened on my large bi-amped JBL 4430 loudspeakers at a normal volume level. Since I had two files open at once it was easy to play them in rapid succession with no delay. I played a few seconds of one file, then the same few seconds of the other, then the first again, and so forth repeatedly.
We played several different short sections of Grekim's tune this way, and neither of us heard any difference between truncated and dithered. Just for fun I found a spot where the ending chord had almost completely faded out, raised the playback volume about 20 to 30 dB, and again played the same section in both files. Even with the volume cranked on a very soft passage we still heard no difference. Even I was a little surprised by that!
One of the claimed features of using an outboard word clock is improved sound quality due to lower jitter. So to test this we played the same sections in Grekim's mix through my Delta 66 sound card using either its internal clock or the clock in his Apogee 8000. We also tried my SB Live sound card using its internal clock. With this test there was no way to switch instantly because I had to stop playback and change Sound Forge's playback device, and Grekim also had to switch the Apogee. The fastest we managed was about five seconds, and Grekim and I both agreed it was tough to tell if anything really changed. Not that I expected or heard any change myself.
At one point Grekim thought he heard a change in the stereo image. That part of the tune sounded to me to have a stereo widener effect on one of the instruments, and I heard the same "out of phase" type sound with all three outputs. After going back and forth a few more times Grekim conceded he was not able to identify with certainty any difference between the three output devices.
SOUNDBLASTER VERSUS APOGEE
For this test we recorded three different instruments, one after the other, simultaneously through my $25 SoundBlaster and Grekim's $6,000 Apogee 8000. Sonar allows recording from multiple sound cards at once, so I used a splitter to send the output of my Mackie 1202 mixer's preamp to the SoundBlaster's line input and also to the Apogee's line input.
I didn't want to apply any volume changes or other processing after the fact, so I played a 700 Hz test tone through my loudspeakers to match the record levels as closely as possible by eye using Sonar's record meters. I believe we matched the levels to within at least 1 dB, and audibly both tracks sounded the same volume. I recorded Grekim playing my Yamaha acoustic guitar and a single triangle hit using my DPA 4090 microphone. We also recorded a few clave hits which I played.
A triangle is very demanding of A/D conversion because it contains frequencies well past 20 KHz, and also contains many different high frequencies. Unlike string and wind instruments, triangles (and cymbals and bells) contain many overtones that are not necessarily related to the fundamental and can also occur closer in frequency. Using Sound Forge's spectrum analyzer I confirmed that the triangle we recorded had substantial energy at 16 different frequencies between 10 KHz and 21.1 KHz. The highest three overtones were at 20,055 Hz, 20,535, and 21,095 Hz. Therefore, this is a demanding test because the A/D converter's filter needs to avoid audible aliasing and IM products caused by those nearby high frequencies.
Sonar's linked Solo feature makes it very easy to switch playback between the SoundBlaster and Apogee recordings. I enabled Solo on only one track, then linked the Solo buttons on both tracks. So clicking either Solo button toggled playback between one track and the other instantly with no gap. Grekim listened with his expensive headphones, and was unable to hear any difference on any of the three instruments.
Some people may not like what we found, but that's not my fault. Neither of us heard any difference between dithered and truncated even with very soft material played back at an unnaturally high volume. Nor did we hear any difference between the jitter in my SoundBlaster card, my Delta 66, or Grekim's Apogee. Nor did we hear any difference in the three instruments recorded through my SoundBlaster and his Apogee.
One failing of this test is we had a bit of low-level hum because my Mackie mixer's preamp direct output is unbalanced. The hum is well below the signal and not audible at normal levels on my SoundBlaster, but it's slightly audible on the tracks we recorded through the Apogee.
I didn't hide the file names to make a blind test, so here they are and you can listen for yourselves to draw your own conclusions. All of these files are mono, 24 bits at 44.1 KHz.
Guitar into SoundBlaster
Guitar into Apogee
Claves into SoundBlaster
Claves into Apogee
Triangle into SoundBlaster
Triangle into Apogee
Here's a final thought about critical listening comparisons: It's not so much that we need to do tests like this blind, but switching quickly is key to knowing if the sound really changed. That is, as long as people are being intellectually honest, there's no need to hide from them what they're hearing. However, the A and B versions must be presented in a way that makes it easy to hear if there really is a difference. It seems to me the key to this, besides carefully matching volume levels, is being able to switch instantly back and forth in mid-stream without clicks or drop-outs. Sonar's linked Solo buttons makes this very easy, and I've used that many times to audition stuff like this in the past. It's only a little more difficult to switch quickly in Sound Forge. You load both files and put the cursor in the same place for each. Then clicking either file's Play button stops the other one and starts the current one. Each time you click Play the selected file begins at the same place.
Ethan Winer has been an audio pro and skeptic for most of his adult life. He now heads up RealTraps, where he designs acoustic treatment products for recording studios and home listening rooms.
Entire contents of this web site Copyright © 1997- by Ethan Winer. All rights reserved.