Audio Research Reference CD9 CD player/DAC Measurements
To measure the Audio Research Reference CD9, I used Stereophile's loan sample of the top-of-the-line Audio Precision SYS2722 system (see www.ap.com and the January 2008 "As We See It"); for some tests, I also used my vintage Audio Precision System One Dual Domain. To test the CD9's performance via USB, I used my MacBook Pro, running on battery power, with ARC's supplied USB2.0 driver (v.2.03.14), and played test-signal files with Bias Peak Pro, using Apple's AudioMIDI utility to ensure that the sample rate and bit depth were correctly set for each file.
Looking first at the CD9's behavior as a CD player, it offered superb error correction. Tested with the Pierre Verany Test CD, which has laser-cut gaps of varying lengths in its data spiral, the CD9 played track 33 (1.5mm gaps) with just one glitch, then occasionally glitched on tracks 35 (2mm gaps) and 36 (2.4mm gaps). It didn't have real trouble until the gaps were 3mm long!
The CD9's maximum output level at 1kHz was 5V from the balanced XLR jacks and half that from the single-ended RCAs, as expected. Both sets of outputs were non-inverting; ie, they preserved absolute polarity. (The XLRs are wired with pin 2 hot.) The output impedance at high and middle frequencies was 620 ohms from the balanced outputs, 307 ohms from the single-ended. At 20Hz, these figures respectively rose to 1400 and 604 ohms, which suggests that the CD9 should be used with preamplifiers having at least a 15k ohm input impedance.
The S/PDIF and AES/EBU inputs successfully locked to datastreams with sample rates of up to 192kHz. The TosLink input didn't work above 96kHz, which is normal for this type of connection. The USB connection worked with data having sample rates up to 192kHz, including 88.2 and 176.4kHz. Apple's USB Prober utility identified the CD9 as "AUDIO RESEARCH CORP DAC," but gave no information about the timing-synchronization protocol, this being handled by the "vendor-specific" driver program.
The CD9 offers four operational modes: Fast and Slow reconstruction filters, each with the option of being upsampled to 176.4kHz with 44.1 and 88.2kHz data, and to 192kHz with 48 and 96kHz data. Fig.1 shows the impulse response at 44.1kHz of the Fast filter and no upsampling. The time-symmetrical ringing reveals it to be a linear-phase FIR type, and wideband analysis of white noise sampled at 44.1kHz (a test suggested to me by MBL's Jürgen Reis) shows that it rolls off very quickly above 22kHz (fig.2, red trace). The blue trace in fig.2 is a similar analysis done while the CD9 played data representing a full-scale 19.1kHz tone. The Fast filter very effectively suppresses ultrasonic images of the tone, other than one at 69.1kHz (88,20019,100Hz), and the second harmonic can be seen at 74dB (0.02%). The noise floor in this graph looks very dirty.
Figs. 3 and 4 show the impulse response and the wideband spectral analysis of the CD9's Slow filter, again with upsampling disabled. The impulse response has just two well-damped cycles of ringing on each side of the pulsesomething reminiscent of Wadia's DigiMaster reconstruction filter. (Audio Research and Wadia are now both owned by Fine Sounds from Italy, along with Sonus Faber and Sumiko.) However, while the rate of ultrasonic rolloff is much slower than the CD9's Fast filter, which means that the image of the 19.1kHz tone at 25kHz is suppressed by just 12dB (fig.4, blue trace), this filter is actually better behaved in its stopband than is Wadia's filter (see fig.2 here).
To my surprise, while upsampling had no effect on the Fast filter's behavior in either the time or frequency domains, it had a drastic effect on the Slow filter's behavior. The well-behaved impulse response seen in fig.3 now resembled that in fig.1, and the filter's characteristic wideband spectral analysis in fig.4 was now identical to that of the Fast filter (fig.5). If you like the sound of the CD9's Slow filter with CDs, do not use upsampling.
The Fast filter's frequency response, with or without upsampling, rolls off by 0.5dB at 20kHz, with the higher sample rates continuing the smooth rolloff before dropping sharply at half of each sample rate (fig.6). Without upsampling, the response with CD data dropped off in the top octave, to reach 3dB at 20kHz (fig.7). At higher sample rates, the Slow filter offers an ultrasonic rolloff similar to the Fast filter's, but without the sharp cutoff just below half of each sample rate. Channel separation at 1kHz was excellent, at 110dB in both directions (fig.8), but this decreased to 84dB at 20kHz, due to the usual capacitive coupling between channels.
For consistency with my tests of digital products going back more than two decades, my first test of a DAC's resolution is to feed it dithered data representing a 1kHz tone at 90dBFS with 16- and 24-bit word lengths, and sweep a 1/3-octave bandpass filter from 20kHz down to 20Hz. The result is shown in fig.9, with the 16-bit spectrum the top pair of traces and the 24-bit spectrum the middle pair. The increase in bit depth drops the high-frequency noise floor by 12dB, implying resolution of around 18 bits, which is sufficient to allow the CD9 to resolve a dithered tone at 120dB, at least in the left channel (bottom solid trace). However, note the low-frequency peaks in these spectra, which are a little higher in the right channel (dotted traces) than the left. These are due to interference from the power supply, and can also be seen in an FFT-derived spectral analysis with the same data (fig.10). Because these spuriae are at 60Hz and its odd-order harmonics, that strongly suggests that they are due to magnetic interference from the power-supply transformer. The levels of all the spuriae are way too low to be audible; even so, they give rise to the dirty-looking noise floors in figs. 2, 4, and 5. Note the low-level spikes at 5.1 and 8kHz in fig.10, particularly in the left channel (blue trace). These may be idle tones of some kind, but are undoubtedly sonically innocuous.
The spectral spikes representing the tone at 90dB in these two graphs peak at exactly 90dB, implying very low linearity error. This was confirmed by a separate test (not shown). The CD9's reproduction of an undithered tone at precisely 90.31dBFS was excellent (fig.11), the three DC voltage levels described by the digital data being clearly resolved, though the low-frequency noise results in some offset in the two channels. Even at this very low signal level, the CD9 produced a well-formed sinewave with 24-bit data (fig.12).
The CD9 was uncomfortable driving very low impedances. But into an appropriately high impedance, the dominant harmonic distortion in the left channel was the third harmonic, at a low 89dB (0.003%, fig.13, blue trace); and, in the right channel, the second harmonic, at 73dB (0.02%, red trace). Though some higher-order harmonics can be seen, these are all well below 100dB (0.001%). The CD9 also did very well with the demanding high-frequency intermodulation test, although, as expected, the Fast filter (fig.14) was much more effective than the Slow filter (fig.15) at suppressing ultrasonic images of the 19 and 20kHz tones.
The power-supplyrelated spuriae made the spectrum of the CD9's output while it reproduced a 16-bit version of the J-Test signal from CD (fig.16) a little difficult to analyze. However, other than the three sideband pairs of unknown origin at ±1.6kHz, ±1.66kHz and ±1.71kHz, which were absent from otherwise similar spectra taken with 16-bit J-Test data fed to the S/PDIF and USB inputs, it looks as if the CD9 is not accentuating the odd-order harmonics of the LSB-level, low-frequency squarewave. Some spurious tones and a sideband pair at ±1.45kHz, again of unknown origin, can be seen with 24-bit external data (fig.17), particularly in the left channel (blue trace).
Overall, Audio Research's Reference CD9 measures well, though I would steer clear of using its upsampling feature with the Slow filter.John Atkinson