ATF™ Adaptive Time Filtering

Adaptive Time Filtering or ATF is a resampling process that allows audio data to be interpolated at a higher precision than standard techniques used in off-the-shelve sample rate converters. ATF uses an advanced polynomial curve fitting algorithm that closer matches the original audio data, than sample-and-hold, or piece-wise linear estimation interpolators, commonly used today. This advanced technique allows jitter invoked errors in the resampling process to be minimized to a point where they no longer become relevant, and as this process is extremely efficient, ATF allows Q5™ Upsampling to be implemented on low cost DSPs boosting overall system performance without impacting on cost.

Standard interpolators are of two types, sample-and-hold, or piece-wise linear and typically have a maximum output signal sampling frequency limited to 192kHz. The diagram below shows a comparison between normal upsampling and Q5 Upsampling running at 384kHz using ATF in the theoretical case where all clocking is jitter free. In the following diagrams we will show how ATF provides superior performance in the presence of jitter and in particular how it minimizes the impact of resampling jitter.

The first thing to notice between ATF and other interpolators is the error path between output samples. This error path in a sample-and-hold interpolator will have a path that does not follow the original sample data, as show below. In the case of ATF, the error path is determined by the smooth curve fitting algorithm, which recalculates a new path every sample, resulting in a path that closely follows the original audio input data.

Having an error path that closely matches the original data will have two benefits: resampling errors caused by jitter appearing on data clocks and by using arbitrary non integer sample rates e.g. 44.1kHz to 48kHz will be minimized. In the case of Q5 the extremely small residual error that is left is designed to be beyond the dynamic range of 24bit audio which is far superior of that of sample-and-hold or piece-wise linear interpolators.

Performing a frequency domain analysis of the resampling error invoked by data jitter or arbitrary non integer sample ratios we find that ATF has some interesting properties. Firstly since ATF is a non-linear process resampling errors are uncorrelated with amplitude, which is not the case with the sample and hold interpolator, and secondly due to the smooth time domain function of ATF, error harmonics will be greatly reduced.