Here we see a part of our ear. It is the cochlea with the basilar membrane located in the inner ear. Every spot in the cochlea is sensitive to a specific bandwidth. The highest frequencies are registered at the front, the lowest at the end of the cochlea. The basilar membrane is part of the cochlea. It functions as a base for 15'000 to 20'000 hair cells. Each of these cells is connected to a nerve which connects to the brain. This is but a very short summary of the working of the ear. In truth, the ear is infinitely more complex. We however will limit ourselves to the sensory hair cells which each describe a limited range of frequencies in such a way that they all overlap. The sum of it determines the range of our hearing. On average that is between 20Hz and 20'000Hz. With 20'000 hair cells, their bandwidth is very small and thus very selective. This means that for one specific hair cell—excepting the specific frequency it is tuned to—the rest of the signal gets filtered out at 40dB/octave. Our hearing at this level behaves much like a band-pass filter. That's comparable to the filters found in CD players. Implementing filters of this level in a CD player can thus be described as overkill.


Applied testing methods. NOS DACs seem to score badly whenever you look purely at their measurements. This is partially correct too because the signal of a NOS DAC passes fewer filters before it arrives at the output terminal. Introducing a sharp filter at the output of a NOS DAC should improve the test results when it comes to harmonic distortion and noise. As a footnote, these are common tests prescribed and generally agreed upon by the industry.
Basilair-membraan

23uS NOS mini DAC
The problem with these distortion tests is that when measuring the amplitude domain, they only use static signal like sine waves. As soon as an impulse signal is used, the results become vastly different. At left is the transient response of a NOS DAC. The faint rounding at the top is due to the earlier mentioned mild filter at 70kHz. At right we see the exact same impulse now passed through a DAC with oversampling. Because our hearing naturally functions as a strong filter, our brains tend to interpret the signal from a NOS DAC as though it had passed through a FIR filter. This is due to the limited bandwidth of our hearing. Looking at the picture at the right, we can wonder how the eventual result will look if our hearing added another equivalent filter. It is well documented by both musicians and audio authorities that especially percussion instruments suffer from this pre/post ringing effect. It is therefore not unfounded when NOS DACs are claimed to sound the most natural of all the alternatives. Because at the same time the test results of all NOS DACs fall short, the question is raised whether we employ the correct tests to accurately gauge their quality. All measurements are, after all, performed without the benefit of any filter.
standaard overbemonstering

 
Jitter has been a point of interest for quite a while. Reducing jitter to extremely low levels is seen as essential for an improved soundstage. Without starting the discussion about how much jitter can still be perceived, it should be mentioned that there are different types of jitter. Because it remains largely unclear how jitter is measured, a short explanation is useful. The jitter in question is nothing more than a deviation in the time domain. You can compare this with the length of seconds for a clock: sometimes they are 0.9, sometimes 1.1 seconds long. On average, they are about one second even though the individual seconds differ in length. This is comparable to a DAC where music consists of a series of digital samples. The samples enter the DAC chip and are synchronised by the clock. The accuracy of the clock is therefore essential not only over longer periods of time but each pulse should be exactly the same in length. [The FIR filter creates an extra link in the chain and contributes, aside from a strong filter, extra oscillation with impulses.]

At the right we have magnified two clock pulses which jump from a low to a high level. The pulses were observed over a longer period of time, then stacked on top of each other. The bottom clock always jumps at almost the exact same moment whilst the top clock shows more random behaviour. Sometimes its jump is early, sometimes late. This is known as jitter. Aside from the fact that jitter can turn up at many points in the audio chain, the detectability of jitter is key to a DAC's performance. The graph clearly shows easily measurable faults for the upper signal. Suppose each square stood for 100 picoseconds (ps). Now the upper signal shows a deviation of at max 5 squares x 100 picoseconds = 500 picoseconds. The more stable the measuring equipment, the more reliable the results from this test will be. However, even measuring equipment will have inherent flaws which might negatively influence the results. When the results are very important, the measuring equipment becomes as expensive as a midsize car. There are also other techniques for measuring jitter like spectral measurements. A stipulation for a good spectral measurement is that a clearly defined signal must be used to measure with. Because the theoretical contents of a signal generate fairly low jitter rates, the jitter which is directly caused by the hardware will show deviations in the spectrum.
jitter2


klokvorming spectrum
[Results of a jitter test from a spectrum analyzer]. The left image shows two spectra. The left is a signal without jitter, hence visible as a vertical needle. The image to the right shows a signal with a broad base caused directly by jitter. Because jitter is a time-based deviation, the spectrum analyser will show jitter as higher and lower frequencies. This 'swinging' signal causes the broad base and shows that a certain amount of jitter is present. It is difficult to determine how much, exactly. Julian Dunn († 2003) was an important researcher in this field. He created a method to better map jitter. He designed a stimulus where the contents are chosen such that in and of itself it should not create any jitter. Because the DAC has to reproduce this special signal, the effects of jitter caused by the hardware will be immediately visible in the spectral measurement. The method Julian Dunn developed is called the Jtest. The image below shows the Jtest applied to a DAC. In an ideal situation, the result would just show the vertical needle. Because of jitter, many smaller needles appear on both sides of the main needle. By measuring the height (amplitude) of these little needles, we can calculate the amount of jitter.


As mentioned earlier, NOS DACs don’t test well. The Jtest is no different. The results of many Jtests for NOS DACs made under similar conditions can be found on the net. Similar conditions are key if comparisons between products are made. Miller Labs are often tasked to perform tests for reviews. These are always performed in the same way and under similar circumstances. It is remarkable that every single NOS DAC performs poorly with the Jtest. Is every NOS DAC designer using inferior techniques? Or are measurement errors the problem? The next test was performed in our own lab. We regularly investigate the characteristics of DAC chips that could potentially be used for hifi. The research tends to be highly experimental and multiple components are connected and replaced by others. This type of research is often aimed at linearity, distortion, switching noise and sound quality. We mentioned earlier that relatively swift electronics tend to sound much better than their more sluggish counterparts. This can be observed with current/voltage converters and other components. Surprisingly, Jtest results turn out far worse when components become faster, with deviations as large as 2 nanoseconds. When we measure jitter in the time domain however, we can see that in the earlier mentioned test the results stay <30ps. Our hypothesis that the result might be skewed by components in the signal from outside the audio band was eventually confirmed. When utilizing a current/voltage converter having a very limited open-loop bandwidth hence low slew rate, the results drastically changed. The Jtest now gave almost the same result as the time domain test. When listening to the music however, the soundstage had turned flat and lacklustre.
Jtest

Jtest op NOS dac   
When using a spectrum analyser for jitter tests, it is striking that not just time but also amplitude deviations lead to a broadened base. Even signals from outside the audio range or a 50Hz hum caused by an inferior power supply will affect these results. So an inaudible 88kHz signal and an equally inaudible 91kHz signal can create a difference signal of 3kHz which in turn modulates the measurement signal. That is why the Jtest can only be used on a DAC which utilizes sharp filters. At left is a Jtest result for a NOS DAC. The needles on both sides of the main signal show the jitter. By current standards, these values are far too high but because they are caused by a lack of sharp filtering, the result is negatively skewed. The same DAC shows a time domain error of only 20ps. This isn't at all reflected in this graph. Because the basilar membrane behaves as a sharp hearing filter, in a certain way it becomes part of the DAC. Measurements on a NOS DAC are therefore not taken at its logical end point but before the filter of our hearing. Herein lies the rub of comparing the actual time domain results and the way in which the Jtest interprets and shows its results.

Conclusion.
NOS DACs have been gaining in popularity for the past few years mostly based on subjective listening tests. Especially people who regularly experience live music appear to have a strong preference for this type of DAC. As Kusunoki mentioned in his article, it is primarily the behaviour in the time domain which gives oversampling DACs their 'unnatural' quality. This shows in the way that percussive instruments sound lacklustre; and in a sort of 'excessive detailing' which causes certain instruments to lose their timbre and natural warmth. The question whether we should follow our ears or the common test results remains on the table. The development of digital audio systems has not yet reached its zenith. We will certainly be confronted with new developments in the future. Certain is that due to high-definition recordings, the need for oversampling and sharp filters has lessened. How to approach the massive variety of CDs with their low sample rate of 44.1kHz remains the question. To oversample or not? Non-oversampling seems to be the preference of musicians and audio professionals. Let your ears decide!


This wrapped up the Cees paper. To fill out a few more tech bits, each of Metrum's proprietary Transient chips is roughly equal to four of the industrial-grade chips used in the Hex. This doubles the Pavane's processing power for true 24-bit dynamic range. Mated to that is 120dB of channel separation from true dual-mono fully balanced signal processing. The I/V converter operates with an extremely fast 1'500V/μSec slew rate and a bandwidth of 500MHz. The OEM M2Tech XMOS USB3 module accommodates up to 384kHz sample rates. The Pavane thus supports original 352.8kHz DXD files and DSD128 converted to PCM in player software like PureMusic. The Lundahl output transformers are XLR-to-RCA summing devices. The performance of the single-ended outputs was carefully matched to the XLRs but given that the Pavane is a true balanced deck, the most direct and pure way to tap its signal is via the 200Ω XLR sockets. Those produce 4Vrms, i.e. twice that of the Pavane's industry-standard 100Ω RCA.