"I have heard the future of audio...and it is digital." Robert Harley on Noise/Mask Ratios
The Audio Engineering Society held its 95th Convention in New York's Jacob Javits Convention Center in October 1993. The gathering is both a trade show for introducing new recording products and a forum for audio scientists presenting papers describing their research. Stereophile has found AES Conventions essential to keeping up with what's happening at the cutting edge of audio technology. Consequently, John Atkinson and I spent four days at the New York Convention, with Larry Archibald dropping by for a few papers and exhibits. Peter Mitchell attended as well, to present a paper on the relationship between musical signals and power-amplifier dynamic requirements.
You can get a good idea of where audio engineering is headed just by looking at the paper and workshop titles (footnote 1). This year the Convention was dominated by three areas: Digital Signal Processing (DSP), multi-channel sound, and low-bit-rate coders (PASC and ATRAC, used in DCC and MD respectively, are examples of low-bit-rate coders). A revolutionary method of measuring audio equipment was unveiled during a workshop at this Convention. The measurement system demonstrated measures an audio component's distortion, then predicts the audibility of the distortion by subjecting the measured data to a model of human hearing. A real-time color display shows the level of audio errors in relation to the theoretical audibility of those errors.
The audio component's error levels are measured and compared to the masking threshold (derived from the human hearing model), producing an NMR (Noise to Mask Ratio) in 27 critical bands. The NMR reveals how close to audibility the component's distortions are. In the hardware implementation, 27 separate 1024-point FFTs are performed many times per second. The NMR is displayed in each of 27 bands continuously in real time. The display was easily and immediately interpreted, with the display changing color whenever the noise exceeded the masking threshold.
This measurement system was developed primarily for evaluating low-bit-rate coders, but the technique could be applied to any piece of audio equipment. For example, I measured a much higher level of intermodulation distortion from the Sonic Frontiers SFD-2 digital processor from its unbalanced outputs compared to the balanced outputs. I also heard a more liquid presentation from the balanced outputs. It is sheer speculation—although intuitive—to suggest that the higher level of IM products was responsible for the difference in sound quality. Using the NMR technique, however, it may be possible to objectively quantify such differences between components. The prospect is revolutionary (footnote 2).
An example of how much more powerful the NMR technique is compared to conventional testing was dramatically illustrated by the so-called "13dB Miracle." A short selection of linearly coded music was played as a reference, then played again with its signal/noise ratio degraded to only 13.6dB. The sound was grossly distorted. The musical section was played a third time, again with measured S/N of 13.6dB, but this time the noise was virtually inaudible.
The difference in sound quality between these signals—each having a measured S/N of 13.6dB—was vast. In the good-sounding signal, the noise was spectrally distributed so as to be masked by the correctly coded music signal. The noise constantly shifted in frequency in relation to the music's spectral balance and amplitude so that it remained hidden by the music. The poor-sounding signal had exactly the same amount of noise, but the noise was flat in amplitude and static in level.
Traditional testing would show these two signals as having identical S/N ratios—despite their vastly different subjective noise levels. The two 13.6dB S/N signals were then analyzed with the NMR technique. It was immediately obvious that the poor-sounding 13.6dB S/N signal had noise well above the masking threshold, while the good-sounding 13.6dB S/N signal's noise was hidden beneath the masking threshold—ie, was inaudible. This was a remarkable demonstration, and one that got JA and me excited about the prospects of applying such perceptual measurements to high-end audio products. It's ironic that research into low-bit-rate coding has produced a measurement method that may one day reveal differences between high-end audio components that are not uncovered by traditional techniques.—Robert Harley
Footnote 1: An audio cassette of this and other AES workshops is available from Conference Copy at (717) 775-0580. In addition, copies of individual papers can be ordered from the Audio Engineering Society at (800) 541-7299.—Robert Harley
Footnote 2: A pioneer in these techniques is Meridian's Bob Stuart. His seminal papers, "Estimating the Significance of Errors in Audio Systems," "Predicting the Audibility, Detectability and Loudness of Errors in Audio Systems," and "Noise: Methods for Estimating Detectability and Threshold," are frequently cited as fundamental works in this area. Bob is not only a designer of great-sounding products, he's also a research scientist working at the cutting edge of correlating measurements with sound quality. See the discussion of these papers in Stereophile, Vol.15 No.1, p.71, and my interview with him in Vol.14 No.9.—Robert Harley