Digital Processor Reviews

Before I discuss the results of my auditioning, I'll present the measured results. These were all taken in the digital domain using an Audio Precision System One fitted with the DSP option. The Meridian 518 was fed a data signal from the AP's AES/EBU output, set to a 20-bit word length. The 518's AES/EBU data output was looped back to the AP's digital input.

In the digital domain, the data word length is inexorably tied to the ultimate level of the noise floor. The longer the word, the lower the noise, with an approximate increase of 6dB in the system S/N ratio every time the word length is increased by one bit. Another way of looking at this is to realize that a linear-PCM system is deaf below a level related to the word length. Thus in an 8-bit system, like early computer soundcards, nothing below -48dBFS can be encoded—"FS" stands for Full Scale"; a 0dBFS signal exercises the full dynamic-range capability of a digital system—and the maximum S/N ratio is also 48dB. With the 16-bit words stored on a CD, the maximum S/N ratio is 96dB; and without the use of record dither (discussed at length in this magazine over the past 10 years), the system goes deaf below -96dBFS.

This is shown graphically in figs.1 and 2, the waveform and FFT-derived audioband spectrum of an undithered 1kHz tone recorded at a level of -90.31dBFS: the resultant waveform can be seen to toggle among just three levels, 0 ("digital black") and ±1 LSB. As a result, as well as showing a component at the fundamental level, the spectrum features harmonic distortion components at 2kHz, 3kHz, 4kHz, 5kHz, 8kHz, and 9kHz. The individual noise components in the fig.2 spectrum can each be seen to lie around the -120dBFS mark: if added in a Root-Mean-Square (RMS) manner, their sum would lie at the -96dBFS level, confirming that this is a 16-bit system.

Now look at fig.3: this is the same kind of FFT-derived spectrum, but now with 20-bit data words representing the -90.3dBFS 1kHz tone. The entire noise floor can be seen to have dropped significantly, with the individual components now lying below -140dBFS and their RMS sum at -120dBFS. The magic worked by the 518 is revealed in fig.4. This is a spectrum derived from 16-bit data, yet over almost all the audioband the noise components are as low as with true 20-bit data. Although the data word length is restricted to 16 bits, as it would be on a CD or DAT, the signal resolution is closer to 20 bits!

Are we getting something for nothing? Of course not. Look at the right-hand side of fig.4. Above 4kHz the noise rises with frequency, slowly at first but with increasing rapidity until, at the 22kHz band edge, each component is on average as high in level as the fundamental tone. What the 518 has done is "shape" the noise floor, pushing as much of it as possible to higher frequencies, where the ear is less sensitive, and dropping it in the midrange and treble regions, where the ear is most sensitive.

This can be seen in figs.5 and 6. For reference, fig.5 shows the waveform of a 1kHz tone at -90.3dBFS encoded with 20-bit precision. By contrast, fig.6, plotted to the same vertical scale, shows the waveform produced by the Meridian 518 when processing digital silence: far from there being silence, there is a considerable amount of high-frequency noise present, its peak-peak amplitude several times that of the reference sinewave in fig.5. The overall RMS level of this "blue-colored" noise is high, around -60dBFS. But as I said before, all this energy lies in a region where the human ear is very insensitive. And if you can't hear the high-frequency noise, it might as well not be there!

Note that this increase in effective resolution does not depend on the nature of the input signal. The source may have acoustic or electrical noise well above the 16-bit noise floor, but the 518 can't distinguish noise from music: everything in the source data is processed as though it were the wanted signal. On the master of Stereophile's Concert recording, for example, the analog tape hiss lay around the -60dBFS level. The difference between the use of noise-shaping (via the 518's predecessor, the Meridian 618) and straight truncation from 20-bit to 16-bit data was clearly audible, however, even in what turned out to be a double-blind test (see "As We See It," February 1995, Vol.18 No.2, p.3).

What it does depend on is the performance of the D/A converter downstream of the Meridian. But if your D/A is capable of true 20-bit resolution—and the measurements in this magazine's reviews are a good source of this information—you can perform the apparently impossible, as is shown in fig.7. To produce this graph, I fed the 518 with 20-bit data representing a 1kHz tone at -90.31dBFS, noise-shaped those data to 16-bit precision using the "B" algorithm, but then used the Meridian's digital gain-adjust function to reduce the signal level by 24dB. The result is a 16-bit datastream, nominally deaf to anything below -96dBFS, that carries a sinewave encoded with a peak amplitude of just -114dBFS. Wacky, wonderful stuff, eh?

But remember, this magic depends on the shaped noise not being audible—which depends on the playback level being correct. Slap in gobs of analog gain after your D/A and you'll start to hear the shaped noise—and it's not pretty. But as long as the recording engineer has chosen the optimum noise-shaping algorithm and you're playing back the 518-processed digital recording at a playback level around what was intended, you're home free.