Vulnerability of the signal

Appendix 2 – Hints on Mastering	46

Appendix 2 – Hints on Mastering with noise-shaping

There are a few difficulties associated with employing noise-shaping schemes to maximise the subjective dynamic-range of a digital audio channel. These are now reviewed.

The available dynamic range in a channel is only as wide as the narrowest step. No matter how much care is taken with a recording, or to maintain the maximum word-size right up to a noise-shaped dithered truncation to e.g. 16 bits, the whole benefit can be thrown away if at that stage a non-dithered signal process is carried out on the material. This is not unheard of, a mastering house may e.g. choose to pass the signal through an undithered dc filter!

This caution simply says that every process on the signal leading up to delivery to the customer has to be reviewed for its impact on the noise- floor and transparency. Such errors will be cumulative, but just one will destroy any dynamic-range enhancements.

Audible results of digital errors and editing

Normal digital channels have provision for error detection and correction. In sophisticated apparatus, short bursts of erroneous samples are interpolated to reduce the audible effect. It is becoming more common in CD replay systems, that the only error handling is a first-order hold; that is when the system detects an erroneous sample, the previous valid sample is output in its place.

This trend has partly been a response to the fact that errors on CD tend to be single missing samples at an acceptably low rate. There are however marked variations in this performance between players.

On normal music this strategy results in a minor modification to the waveform and it usually passes unnoticed. The introduced sample is of similar size to the signal, and is therefore readily incorporated or masked. On normal material we also notice that errors occurring in silent passages result in near-zero additions which tend to be inaudible.

Let us now consider the circumstance where a signal has been passed through a noise-shaper with a function like Shape D. In loud passages errors will be masked just as the normal circumstance. However channel errors occurring in quiet or silent passages will tend to result in an audible click (since the HF noise is boosted) with little or no material to mask it.

A similar problem pertains to editing ‘shaped’ material in quiet passages. The requirement exists to maintain apparent continuity in the shaped noise. We have observed clicks due to replay errors when shaped material is re-played on cheap CD players, but no difficulty at all on reasonable or up-market units.

Requirements for D/A conversion

Noise-shaping can be used to extend the audible dynamic range of a channel beyond that which would be expected for a normal digital channel of the same word-size. The essential principle by which this is achieved is representing numbers below the LSB as an average of successive bracketing numbers. For this system to work, the channel must exhibit differential linearity equal to the resolution required.

We can illustrate this best by example. Let us assume that a signal has been encoded with 20-bit precision and that a shaper type D has been used to convey this signal onto a 16 bit CD channel. Measurement in the digital domain shows very clearly the success with which the subjective

518 User Guide