Truncation vs. Dither?
May the best method win, by JD Mars

 

The mere mention of "truncation" seems to instill fear in the hearts of audio men and women. Dither sounds a bit more genteel, but I can't seem to get it out of my head that it rhymes with "whatever," and of course spoken with a valley accent. Oh well.

That doesn't undermine the seriousness of a discussion of truncation vs. dither. Oh no.
But after all is said and done, after you slice it and/or dice it and the smoke clears, you're going to be left with a decision to make that is mostly subjective in nature. Sorry about that.

The Playing Field

When an audio file or digital recording is created with a particular bit resolution and is transferred to another device or application with a lower bit resolution, this change in bit width must be resolved in some way. Do nothing, and the receiving device will truncate, or chop off the bits that it can not handle. In other words, if your recording has been made at 24 bits and you're transferring to a device that can only handle 16 bits, the "extra" 8 bits get chopped off and are lost in oblivion. Gone. Kaput. Buh bye.


Only if your audio device or software has an option to add dither can the decision to choose dither be made. You may be able to add dither after the fact, alleviating the need to make a permanent irreversible decision on the spot. Either way, the 8 dits are lost. Dither is a way of compensating for the truncation's inevitable increase in audible quantization error before the truncation occurs.

A Truncation Situation

The implication? Bit depth represents dynamic range in digital recordings. A recording that is done in 24 bits will have a greater dynamic range than a recording done in 16 bits. Dynamic range is the difference, from softest to loudest, that a recording device is capable of. Dynamic range specs will include system noise, so that signal to noise ratio will always be less than dynamic range. In other words, s/n is a subset of dynamic range.


Analog systems have some inherent system noise. Digital systems introduce no system noise, so where the noise factor enters in is with something we call quantization error. Quantization, simply put, is the amplitude component of the digitized audio. Quantization error occurs when the program signal becomes so quiet that the lowest, or "least significant bit," can no longer adequately process the audio. There is dynamic range between the 0 and the 1, but 0 and 1 can't express that -- it's one or the other (no pun intended). The resultant square wave is the equivalent of a digital noise floor, which in effect is a random array of "artifacts", often expressed as "granulation noise," since the perception of these digital artifacts is that the audio has attained a grainy quality. Understand that this happens at the quietest level -- when you make your final Truncation vs. Dither decision, you're going to have to listen with your critical ears.

In terms of digital noise, the greater the bit depth, the quieter the digital noise. Each added bit reduces the audible level of quantization error by 6dB, since each bit adds 6dB to the dynamic range. This is a substantial reduction in perceived noise when the bit depth is maintained throughout the recording, but if audio CD is the ultimate goal, 16 bits is the limit in the digital domain.

As stated, truncation is inevitable in a transfer to a lower bit depth. The real question is not so much truncation vs. dither, but rather, doing nothing vs. adding dither.

To Dither or Not to Dither

Quantization error is a digital reality, and also applies to the analog to digital conversion of audio. Since the answer to "what came first, the A/D or the D/D?" is "A/D," dither was initially an analog white noise that was added to the signal to displace the audio signal level above the least significant bit. So we're actually adding analog noise to prevent digital noise? Yup, that's right.

The Fletcher-Munson curves show higher responsiveness to sounds in the 3 kHz to 4 kHz area. Intensity in level varies while the ear perceives equal volume at all frequencies (Click image for closer view).

As insane as that might sound, you might consider this analog noise to be a more known quantity that those nasty random artifacts, and in that sense more desirable. A little system noise or tape hiss is something that our ears have been used to for years, and perhaps far more digestible than the unknown. Also, quantization error is considered distortion, whereas a noise is just a noise (but is a sigh just a sigh?). In the realm of digital to digital, this analog noise is created or synthesized using an algorithm, generally specific to the device or software.

There are a few permutations to this dither signal. "Noise shaping" adds an EQ filter to the white noise signal so that the noise is shaped according to the frequencies that the ear is most sensitive to (as in the Fletcher-Munson curve, or some more modern variation). This noise shaping can also be offered by some systems without dither, and rather is calculated as a function of the errors introduced by truncation.

In Conclusion

If you have the time and the luxury, try transferring your digital audio first by merely truncating. Then, if your audio source device provides these options, transfer using dither, dither with noise shaping, and noise shaping without dither. Listen. Is there a result that sounds most desireable to you?


This may be the bottom line. There are other considerations, however, that would influence your decision not only whether or not you should, but when to do so. Quantization error occurs at the softest or quietest points in the recorded signal, so the content of this recorded signal is something to consider. If are no quiet moments in the program, there may be no audible artifacts until the end of the fade-out, for example.

Also, if your final objective is to pre-master your CD with a variety of cuts, cross fading two cuts that already have dither added can have a double-dither effect (to coin a phrase). Adding dither to the final edited version would be a better choice. Noise shaping may be undesireable if your final product is a sound that may wind up in a sampler, subject to pitch shifting. The audio quality of the noise shaping would then be subject to this pitch shifting. Lastly, noise shaping without dither, which functions in relation to the quantization error, can have varying results dependent on the dynamics of the program material. A "pumping" effect may result from unusual modulation in the quantization.

The choices are yours. Hopefully, the choices that you now make are a "bit" more informed. Pun intended.


JD Mars is the Producer of DigitalProSound.com