iFi audio - The GTO filter
Part 4/4 - Introducing the iFi GTO™ Digital Filter
Transients and Digital Filters
Digital filters introduce time-domain distortions that are unavoidable and a result of the inherent functional principle of digital filters. Digital filters are formed from delay-lines with so-called ‘taps’ usually at one audio sample intervals. Summing the delayed signals with varying gain, positive and negative, forms the filter function. Time-domain distortion, just as amplitude domain distortion is accumulated along the recording and playback chain and cannot be removed unless its exact nature is known.
Among others, the mathematical consequences of this filtering have been described by J. Willard Gibbs after whom the Gibbs phenomenon is named. A square-wave filtered perfectly with a filter that allows only the fundamental to pass, will be a sinewave. Transient distortion, however, will be very high. If a square-wave is filtered but not with a perfect filter that only allows the fundamental to pass, ringing is observed. No filtering restores the perfect square-wave.
Optimising Digital Filters for transients
Considering the responses of the human ear, a digital filter can claim to offer a transient response distortion that does not materially alter the perception of music should have a transient response that has been completed in under 700μs and be free of pre-ringing.
Such a filter could be an asymmetrical FIR filter with no more than 32 taps at 44.1/48kHz sample rate and no more than 64 taps at 88.2/96kHz sample rate, which is precisely what iFi have implemented as GTO Filter.
Given the ‘Haas Window’ within which reverb or ringing is integrated into the main arrival is agreed to be at minimum 5ms (some debate remains to the actual number), one may expect filters that require more than 5ms to complete their impulse response (or have in excess of around 220 Taps at 44.1kHz sample rate) to have the potential to be significantly audible as alteration.
By contrast a 5μs transient resolving ability of the ear would require a sample rate of around 200kHz to ensure a transient response, while a 20μs transient resolving ability of the ear would require a sample rate of over 50kHz, with a good middle ground represented by a 88.2/96kHz sample rate.
The above suggests that at 44.1kHz sample rate digital audio lacks time domain resolution to be transparent and the common ‘long’ symmetrical FIR digital filter will have the potential to cause audible problems. It is thus desirable to use a higher sample rate and ‘short’ asymmetrical digital filters to ensure recording and playback are audibly transparent.
At this point in time even higher sample rate recordings tend to utilise “long” symmetrical FIR Filters, so their benefit is diluted. That said, if we know the filter response applied at the time of recording, the time-domain distortion in the recording could in theory be reversed and replaced by one that is not audible, this action is asserted for the complete end to end MQA coding system promoted by MQA Ltd, which is now supported by all current iFi products.
For recordings not available in MQA derived from analogue Masters or genuine high-resolution digital masters with a known provenance, selecting iFi’s GTO filter presents the highest possible fidelity.
How iFi has implemented the GTO™ Digital Filter
A digital filter must be implemented somewhere, perhaps on an FPGA like the Xilinx Spartan 6 as used in the Pro iDSD, a DSP Chip like an Analogue Devices Sharc or Blackfin. Recent developments on the XMOS Platform have led to our becoming very familiar with offerings from the new XCore200 and the onboard DSP capabilities baked-in.
In the Pro iDSD we choose to use the XILINX Spartan 6 FPGA which also performs the DSD 1024 remastering, simply because the 16-Core XMOS is substantially loaded already.
Using any option requiring dedicated hardware would mean this filter can only be made available on products in development at higher price-points as FPGAs or DSP chips sufficiently capable are costly or on products using the xCore-200, which is on-board of the Pro iDSD.
GTO Trickle-down
Unlike the xCore-200 series which include built-in DSP capabilities, implementing MQA on xCore-100 required the ground-up design of a DSP engine. With the DSP engine now available, our natural inclination was to have our new filter baked onto the XMOS Firmware.
Technical assistance came from an unexpected source; MQA. Using very different filter parameters as part of a more complex system including the decoder section requires serious DSP chops. The implementation of MQA rendering on the original xCore-100 series XMOS processors widely used in the iFi products required extensive DSP work.
This took some extra work, and some fine-tuning of resource allocation, however with the version 5.3C of our XMOS firmware we now offer this new, Global Transient Optimised digital filter to all our customers, including just about all our legacy products, as an alternate option to the filters implemented in the DAC chip that apply for version 5.30 and before.
Technical backdrop
The ‘Gibbs Transient Optimised’ filter was developed by iFi according to our specifications in conjunction with the MQA team. We must make clear that GTO is not directly related to filter types used by MQA, it is not “MQA through the backdoor”, but instead what we feel is the optimum solution for the playback of digital audio that has not undergone the MQA process. We would like to thank MQA for their technical assistance in integrating this into our firmware.