"I have heard the future of audio...and it is digital." Peter W. Mitchell on Noise/Mask Ratios
Until now, developers and reviewers of perceptual coders have relied only on listening tests, because conventional measurements provide little guidance. At the New York AES convention two new methods were demonstrated that can provide valuable information about audible flaws in low-bit-rate coders. Both operate on the same principle as the perceptual coders themselves—namely, the concept of "masking" in critical bands (see later).
The cochlea of the inner ear behaves as a bio-mechanical spectrum analyzer. A tone at any frequency can be masked (made inaudible) by louder sounds at adjacent frequencies. A "critical band" is the range of frequencies, approximately a third of an octave wide, over which such masking occurs. This has been well known for decades. Conversely, a sound at any frequency will mask (cover up) any weaker sounds at adjacent frequencies. For example, in the presence of a 1000Hz tone at 80dB spl, a tone at 900Hz or 1200Hz will be inaudible unless it is louder than 50dB. (For more details, see our April 1992 report on the technology behind DCC.)
Traditional audio measurements have employed a test tone at just one or two frequencies at a time, while an analyzer measures the strength of the resulting harmonic or intermodulation distortion components. In perceptual coders the audible frequency range is divided into 30 or more sub-bands, and the available bits are divided among these according to the strength of the signal in each band. If a test signal contains just one or two frequencies, the coder can allocate all of the bits to just the one or two sub-bands that contain those frequencies, coding them with full 16-bit precision. The usual result is that perceptual coders produce virtually perfect measurements with conventional test signals; the test tells you nothing at all about whether a low-bit-rate coder alters the sound.
During the AES convention's workshop on perceptual coding, Fraunhofer engineers blew the socks off most attendees by demonstrating an amazing device: an NMR analyzer. The impact of NMR measurement is likely to resemble the bomb that exploded under the World Trade Center last February: The twin structures of audio design and reviewing won't topple, but they will be thoroughly shaken, followed by the construction of a more secure foundation.
NMR stands for "noise/masking ratio." In this context, "noise" is anything added to the signal that doesn't belong there—coding errors, various types of distortion, quantizing noise, et al. It produces a continuous on-screen display showing (in different colors) the signal level at each frequency and the amount of unwanted "noise" around the same frequency. At each instant it also calculates the ratio of the noise to the level of the masking curve at each frequency, and combines these results to obtain the overall noise/masking ratio. The NMR is expressed in decibels; if the level of noise is much lower than the masking threshold, the NMR is a very negative number. If the NMR approaches 0dB or becomes positive, the decoding errors are plainly audible.
The NMR analyzer, which requires several very fast and powerful DSP processors to do its many calculations (meaning that it is likely to make it a very costly system), can measure dynamically varying signals. This is an important advantage: some of the most important flaws in perceptual coders are time-dependent errors. The design of sub-band filters involves a difficult compromise between bandwidth and response speed, which is why perceptual coders often make momentary errors at transient beginnings where the character of the signal is changing rapidly. [See this issue's "Letters" section.—Ed.]
The demonstration compared the performance of a DCC recorder (with an average NMR of -12dB) to a MiniDisc deck (-8dB). It also showed the much lower error content of a prerecorded DCC tape (-20dB), illustrating the superior performance of the professional PASC encoder that is used to master commercial tapes. The on-screen display also showed the very different characteristics of DCC (whose coding imperfections occur mainly at high frequencies where they are unlikely to be heard) and MD (whose coding errors are distributed uniformly across a broad frequency range). It also illustrated the gradually worsening NMR caused by tandem coding (repeatedly passing a signal through DCC coding).
Every audio reviewer in the room salivated at the prospect of using the NMR analyzer, not only to assess the performance of perceptual coders but also to investigate subtle flaws in other products. For example, it might reveal dynamic (signal-dependent) distortions in power amplifiers that measure identically with steady-state signals. JA and RH are excited about the potential of this measuring system to aid in identifying subtle flaws at the margin of audibility. David Ranada, recently appointed Technical Editor of Stereo Review, hopes eventually to transform that magazine's test procedures until they all are perceptually based, ranking products according to the severity of their audible imperfections.
This won't happen overnight, but the future looks exciting. If the price of the NMR measuring system is too steep for individual product designers or magazine reviewers, perhaps one unit could be installed at a central location and shared among many users—the way that the excellent speaker-testing facilities at Canada's National Research Council are available to various speaker designers and reviewers.—Peter W. Mitchell