Digital recording is good, but it's nowhere near the perfection that was first claimed for it, even though on paper you get low distortion and a ruler-flat frequency response across the entire audio band. Part of the problem with conventional linear PCM (Pulse Code Modulation), or sampling, is that steep anti-aliasing filters are required just above 20kHz, and these tend to ring, adding high-frequency energy to the original sound for a significant time after the original transient that set the filters ringing has ended. Pundits tell us that the answer is to use more bits to improve resolution, and to use a higher sampling frequency to offset the need for sharp filters, but there's still a problem -- linear systems divide signals up into equal steps of level in equal packets of time, so whatever artifacts are generated by the sampling process are always correlated. In a well designed system, these artifacts may be of a very low order indeed, but because any side-effects are rigidly related to the sampling frequency and quantisation step size, they invariably concentrate at specific points within the frequency and time domains, making them more likely to be noticed. Furthermore, because low-level signals are represented by fewer bits, the distortion level rises dramatically at low signal levels, and technical subterfuge, such as noise-shaped dither, has to be used to extend dynamic range and low-level clarity. But now, scientist Rila Oplof, working at the Bulgarian Institute of Railway Kinetics (B.I.R.K.), thinks he's found a better way, and oddly enough it's closely related to the railway technology to which he still devotes most of his academic life.
To spread unwanted artifacts evenly in both the frequency and time domains, it's necessary either to have an infinite number of quantisation levels and an infinite sampling frequency (clearly impossible), or to break up the linear relationship that has constrained linear PCM for so long. Instead of a regular sampling frequency, Oplof's system uses an Analogue-to-Digital conversion process based on a pseudo-random number system to generate a clock frequency that seems to vary or 'wobble' at random between 44.1kHz and 48kHz. The pseudo-random number cycle (which was apparently generated by computer from the relationship between the Bulgarian National Railways timetable and the times the trains actually ran) repeats on a two-second cycle, and includes an embedded code for synchronising the D-A conversion system at the other end of the chain. In practice, the audio is clocked out with its original time relationships intact, but artifacts due to the clocking frequency are now smeared gently over a wide area, rendering them inaudible. In many respects, this emulates how analogue artifacts are randomised, but in the case of Wobbly Bits, the artifacts remain at a much lower level.
This leaves the equal quantisation step size problem, and Rila approached this in a very similar way. By changing the resistor ladder in the A-D converter that defines the quantising step size from equal values to pseudo-random values spread over a 2:1 ratio, he found that the analogue signal could be quantised into a series of irregular step sizes, and if the resistor sequence is rotated every clock cycle, each successive sample is quantised differently. This time the random number sequence is based on prime integers (derived from the differences in time between the announcement of certain digital mixing consoles and their actual arrival in the shops? In which case it probably involves imaginary numbers!). The length of the cycle is approximately 1.7326987519353728 seconds, so it remains essentially uncorrelated with the sample frequency loop cycle.
Once again, this process has its equal and opposite counterpart in the output converter, so that the original audio is correctly reconstituted with very low noise and distortion. Quantisation distortion, which most people regard as objectionable, is redistributed as almost perfectly gaussian noise, and if the appropriate pseudo-random quantisation step sizes are chosen, this noise can be shifted into the 15-18kHz part of the audio spectrum, where it is less likely to be audible. This technique completely does away with the need for noise-shaping dither systems (as it achieves the same thing automatically), and actually produces less background noise and distortion, while bringing about a noticeable increase in low-level signal transparency and enhanced stereo imaging. Indeed, panels of listeners thought the Wobbly Bits system sounded far more accurate than the 30ips (inches per second) analogue tape from which it was copied!
Though Wobbly Bits systems (based around Oplof's appropriately-named Wobble Board and to be marketed under the brand-name of Kinetix) have already been tested for recording, there are obviously interfacing problems with other digital equipment -- locking to a rapidly varying clock will throw all normal digital equipment into a tight spin, and attempting to devise DSP mixing and EQ algorithms for variable sample rate, variable quantising step size audio is going to take a lot of lateral thinking. Nevertheless, the only alternative for high-definition audio is to use a vastly high data rate, and given that demands on digital editing and mixing systems are always increasing, this is not a direction most manufacturers want to consider. Indeed, it seems that Wobbly Bits is presently the best hope for low bit-rate, high-quality uncompressed audio, and systems working on this technology are expected to be available by next April.