Sample rates, wordlengths and the DN9848
We are often asked questions such as “why don’t you quote the number of bits for your analogue-to-digital converters (ADCs)?” by people wishing to compare our equipment with products from other manufacturers. This has been a deliberate policy, because of the danger of making "over simple" comparisons between competing units based on numbers of bits or sample rates. In many cases the actual performance may differ substantially from the "apparent quality" based on the numbers in the specification. So, in response to these questions, here is a summary of the DN9848 architecture, with some background on how this can be sensibly compared with competitor products.
DSP sample rate is 48 kHz. This allows us a theoretical 24kHz audio bandwidth, although we only specify 20 Hz to 20 kHz, and we deliberately roll off above 20kHz. In our opinion, bandwidths wider than this are in general undesirable for live sound, as they merely increase the likelihood of HF driver failure without any sonic advantage. Many people over the years have conducted subjective listening tests comparing 96kHz sampled systems to 48kHz systems and found that they sound different. However, this usually involves different analogue stages, different ADCs and DACs, different phase responses and so on, so it is no surprise that they sound different. On the other hand, if a 96 kHz sampled system is built, and then a 20 kHz digital filter is introduced inside the system, we remain convinced that the result is inaudible. This assumes, of course, that the filter is linear phase and has low ripple in the passband (not always the case !). 96kHz sampling also causes problems with the noise performance of low frequency EQ stages (because the differences between adjacent samples are smaller), so a 96 kHz system typically requires a longer wordlength to achieve the same noise performance as a 48 kHz one. The one advantage of a 96kHz system in live sound is that it is possible to reduce the latency (delay) through the system a little. Note also when comparing 96kHz and 48kHz systems that many 96kHz systems specify audio bandwidths of 30kHz or even 40kHz, and then only specify the noise performance up to 20kHz. Clearly if the system is flat to 30kHz, then all the noise up to 30kHz will be arriving at the power amplifiers and should be included in the noise measurement. This is particularly true when oversampling ADCs are used, which have a noise profile that typically rises with frequency.
DSP wordlength is 24-bit, fixed-point (optionally 48-bit fixed-point where necessary for the algorithms). This gives us a theoretical internal dynamic range of 144 dB, so this is comfortably better than the converters that are currently available. Fixed-point versus floating-point is a big discussion, but in general a 24-bit fixed-point system is harder to design than a 24-bit floating-point system but sounds better. This is because when there is a typical loud-ish signal level passing through the unit, the "step size" available between samples is smaller on the fixed-point system. In addition, the step size is fixed, whereas a floating-point system has a variable step size depending on the instantaneous signal level. In other words in a floating-point system the quality of the quiet hi-hat cymbal will be modulated by the signal level of the bass guitar - not generally a good thing... Obviously the floating-point system has a theoretical noise advantage at very low signal levels, but by the time the level is low enough for this to be significant, the ADC and DAC noise will be dominating, not the DSP noise.
The ADC and DAC parts that we use are both "nominal 24-bit" items, but this is essentially meaningless. If a manufacturer claims that they have a "24-bit converter" in their product, then the next question to ask is how you should measure the unit to confirm the 144 dB dynamic range that this implies. In practice no-one is achieving even 20-bit noise performance (=120 dB dynamic range) from a digital system of this kind at the present time. The DN9848 achieves >114 dB dynamic range or "19 bits" overall from input to output. Note that this is an unweighted figure (i.e. flat frequency response). Some manufacturers quote "A-weighted" figures which flatter the unit's performance significantly by applying a psycho-acoustic curve to the measurement. Measurements which specify the dynamic range of the ADC or DAC in isolation should also be treated with caution, since these are often “data sheet” numbers supplied by the IC manufacturer which are rarely if ever achieved in practice. The ultimate safety net is to say “could I verify this measurement myself with an example of the unit and a test set ?” – if you can, then the manufacturer is unlikely to be exaggerating – the potential for embarrassment is too great ! If the figures can only be verified by calculation or internal connections to the circuitry, then the figures may be less useful.
The other key performance issue even for digital products is the analogue audio stages - in particular the difference between bench measurements and real-world performance. KT units are designed to perform not only when connected to test equipment on a bench, but also when driving long cables, unbalanced loads, and in the presence of external electrical and magnetic fields. Issues such as common-mode rejection (especially at high frequencies) and impedance balancing of outputs can have a dramatic effect on the actual performance obtained, as opposed to the "brochure specification".
In the end, the one-sentence summary is "don't worry too much about the bits and sample rates - trust the same real-world performance measurements of noise and distortion that you would apply to analogue".
And after that, there are always your ears...