Ayre Acoustics QA-9 USB A/D converter Measurements
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") to examine the Ayre QA-9's measured performance, analyzing the AES/EBU output data in the digital domain. I repeated some measurements using the USB connection and examined the resultant AIFF files with Adobe Audition's FFT analyzer. I assessed only the QA-9's behavior as an LPCM converter; take it as read that all the measurements I discuss refer to both AES/EBU and USB output formats, there being no difference between them.
The Macintosh USB Prober utility reported the manufacturer string as "Ayre Acoustics," the product string as "Ayre ADC USB Interface," and the serial-number string as "Streamlength(tm)," the latter referring to Gordon Rankin's proprietary asynchronous USB protocol.
The QA9 preserved absolute polarity. Its input impedance was too high for me to reliably measure with my usual voltage-drop method; it is at least 1 megohm at all audio frequencies. With the level control at its maximum, an input signal of 810mV illuminated all the meter LEDs, including the red one, and produced data at 0dBFS. A 728mV signal lit up all the green and amber LEDs and resulted in data with a level of 1dBFS. The first amber LED illuminated at 3dBFS. Each click of the level control reduced the input sensitivity by 2dB. Channel separation (not shown) was >110dB at higher frequencies in both directions, and buried beneath the noise floor in the midrange and below.
Fig.1 shows the QA-9's frequency response with 192kHz-sampled data and the rear-panel DIP switch set to Measure (blue and red traces) and Listen (cyan and magenta). In both conditions the response rolls off smoothly above the audioband, but Measure introduces a sharp cutoff above 58kHz that reaches 10dB at 72kHz. The Listen response reaches 3dB at 70kHz. Fig.2 shows the response with the Listen setting at sample rates of 48kHz (cyan and magenta traces), 96kHz (green and gray), and 192kHz (blue and red). The slow rolloff suggests only a modest degree of anti-aliasing suppression above half of each sample rate. This will not be an issue with 192kHz sampling, given how little energy there is above 96kHz in music, and most likely not with 96kHz sampling. But with things such as cymbals and transient-rich music and LP ticks, the user should experiment with both the Listen and Measure settings if he wants to sample at 44.1 or 48kHz.
All A/D converters have problems with the highest-level signals, which is why it is wise to leave a few dB of headroom when you record. The QA-9 is no exception, though it is better behaved than many other ADCs in this respect. Fig.3, for example, is the digital-domain spectrum of the QA-9's output when fed a 1kHz tone at a level equivalent to 1dBFS. The third harmonic at 3kHz is the highest in level, at 64dB (0.06%), with both even- and odd-order harmonics visible. Dropping the input signal to 6dBFS gives a much cleaner spectrum (fig.4), with the third harmonic now at 74dB (0.02%), while an input signal at 10dBFS both drops the third harmonic by another 10dB, to 80dB (0.01%), and leaves just the second and fifth harmonics visible above the noise floor (fig.5). These three graphs were taken with the input-level control set to its maximum; lowering the control by five clicks (10dB) and increasing the input signal level by the same 10dB to give output data at 10dBFS gave the spectrum shown in fig.6. The third harmonic is still the highest, but has dropped to almost 100dBFS (0.001%), though the second harmonic has risen slightly, particularly in the left channel (blue trace). Despite its zero-feedback analog topology, the QA-9 offers superbly low levels of harmonic distortion.
Intermodulation distortion was also low. Figs.7 and 8 show the digital-domain spectrum of the QA-9's output while it encoded an equal full-scale mix of 19 and 20kHz tones but with the level control set to 6dB, and with the anti-aliasing filter set to Measure and Listen, respectively. The sample rate was 192kHz. Other than the additional rolloff above 50kHz, which reduces the level of the second- and higher-order harmonic distortion and the higher-order intermodulation products, the spectra are basically identical. The 1kHz difference tone lies at just 94dB (0.002%) in the left channel, 106dB (0.0005%) in the right (both ref. the peak amplitude of the input signal), while the highest-level intermodulation products at 18 and 21kHz lie at 74dB (0.02%).
Looking at the QA-9's amplitude error, the level of the output data accurately tracked the level of the analog input signal down to well below 110dBFS. While the QA-9 outputs data with a 24-bit word length, its resolution will be less than 24 bits, of course, due to the converter's self noise. Converter resolution is usually assessed by feeding a 1kHz tone at a level equivalent to 60dBFS and examining the spectrum of the output data. The performance of the QA-9 with this test is shown in fig.9: other than the 1kHz tone and a tiny bump at 21kHz, there is nothing to see other than the converter's noise floor around 146dBFS. This implies true resolution close to 20 bits, which is superb, and almost as good as the dCS 904 converter.
There isn't a straightforward means of estimating the effect of jitter with an A/D converter by looking at the output data. But as the effect of variations in the timing of the master clock will have a greater effect as the signal frequency increases, I fed the QA-9 a 20kHz tone at a level equivalent to 10dBFS, sampling at 192kHz, when any variation in the master-clock frequency will be a proportionately higher percentage of a single clock cycle than it would be at a lower sampling frequency. The resulting spectrum, analyzed in the digital domain, is shown in fig.10: the second, third, and fourth harmonics are all evident, albeit at very low levels. There are spurious tones visible at 2.63, 42.36, and 91.96kHz, none harmonically related to the fundamental tone, but none mathematically related to the sample frequency or forming jitter-related sidebands around the fundamental. I have no idea where these tones come from, but they are all extremely low in level, at 132dB or below.
The QA-9 obviously has no appreciable jitter, though I was intrigued by the slight spreading of the low-level "skirts" to either side of the spectral spike that represents the 20kHz tone. This spreading might be due to very low-frequency, random variations of the clock frequency, but it might also be due to the fact that the FFT is not synchronous (ie, there is not an integer number of signal cycles in the time sample being analyzed), and that the FFT windowing function will have some mathematical limitations. (Depending on which windowing function is used, there will be some leakage from a very high-value FFT bin to the adjacent bins.) This graph was taken with Audio Precision's Equiripple window using a 32,768-points FFT; I did try different FFT windows, including Blackman-Harris, Rife-Vincent 5, and Rectangular, and repeated the analysis with Adobe Audition's FFT analyzer with its Blackman-Harris and Kaiser (180) windows, but with no significant change in the measured spectrum (footnote 1).
For reference, I repeated the test with one of my dCS 904 A/D converters, also sampling at 192kHz and using the AP Equiripple FFT window with the same FFT length (fig.11). While the dCS converter produces similarly low levels of second and third harmonic when compared with the Ayre, its noise floor rises above the audioband due to the noise-shaping used to produce 24-bit LPCM data from its sigma-delta, 5-bit converter. By contrast, the QA-9's ultrasonic noise floor is flat with increasing frequency. The dCS 904 also has the same skirts around the spectral spike, suggesting that this is indeed an artifact of the windowing function, but you can also see that the 20kHz tone now has sidebands, particularly in the left channel (blue trace), though they lie at a still very low 120dB.
Summing up these test results is easy: Ayre Acoustics' QA-9 A/D converter offers superb measured performance that correlates with its equally superb sound quality.John Atkinson
Footnote 1: I say "no significant change," but Audition's Kaiser (180) window, with a 65,536-point FFT, did hint that there was a pair of low-frequency sidebands that were being hidden by the spectral spreading from the other window functions. But at 140dB, these are still inconsequential.John Atkinson