English

High-Resolution Audio

The precision with which audio data can be represented is determined by the word length in an integer format, and to a certain amount by the number of bits in the mantissa in a floating point format. A common consumer format for high-resolution audio data is 24 bit integer.

A high-resolution audio product needs precision management for maintaining the highest possible audio quality throughout the production process.

Once the analogue audio is converted to digital form there remain two types of error sources which should be considered:

• errors in arithmetic due to limited precision when manipulating the audio data itself,
• limited precision due to word length when the manipulated digital samples are requantized for final storage.

Digital Filter Circuits, Extended Precision

In the first case it is necessary to choose proper digital filter circuit structures just to minimize the audibility of quantization noise.  This choice is valid in general for both integer and floating point processing and is particularly true for high-Q filters and swept filters, that is, filters with changing coefficients.

It is also reasonable to use extended precision data format for all calculations, thus using for example some number of low-order bits to keep track of a granular portion of the samples. Obviously this may increase the load of the processing unit, a DSP as a rule, because of the sequential nature of the operations referred to. As an alternative AdiSys has developed a technique based on noise attenuation which performs similar to classical extended precision arithmetic but can be implemented in parallel.

The diagram shows the quantization noise behaviour of a high-Q second-order digital filter implemented with 16 bit integer arithmetic and band-limited white noise as input signal.

Dithered Requantization, Noise Shaping

Since quantization noise tends to be correlated with the audio signal it is important that requantization is performed with minimal audible distortion. With dithered requantization, one can avoid that signal distortion is perceived at the expense of an additive noise source. Dithered requantization is not correlated with the audio signal.

Noise shaping is used to push the majority of the dither noise to frequencies where the ear is least sensitive, and remove noise where the ear is most sensitive.