- #1
- 10,776
- 3,637
Hi Guys
This question came up in the new supposed big thing in audio called MQA (Master Quality Authenticated). Here is a technical overview and a link to its patent:
https://www.soundonsound.com/techniques/mqa-time-domain-accuracy-digital-audio-quality
https://patentscope.wipo.int/search...A34392194DDAEFDDC7.wapp2nA?docId=WO2014108677
They talk about explaining it with triangular sampling and linear interpolation in approximating what was lost in the sampling. This of course is an example of a spline of order 1 - pretty trivial really. But it is claimed better results are obtained using higher order splines:
'Even better results are possible using higher-order ‘B-spline’ kernels, which allow both the position and intensity to be identified of two or more separate pulses occurring within the same sampling period!'
The MQA guys are bit coy about exactly what they do, but the conjecture is they do an an analysis of the music, decide from that what order of spline to use in sampling it down to 96k and put the best way to upsample it in metadata encoded in the bit-stream in some way.
The reason is they claim to want to reduce time smear as much as possible ie when fed a Dirac impulse it has the shortest time response.
OK so far. But in the discussion I am having, some guys claim that above order 1 spline reconstruction is not a FIR (Finite Impulse Response) filter. I know polynomial reconstruction is:
https://www.dsprelated.com/freebooks/pasp/Lagrange_Interpolation.html
I know splines are polynomials between points - but not exactly the same as Lagrange interpolation. However since the whole object of this is to keep the time response as short as possible I can't see how it can be anything but FIR.
Does anyone know the answer - is it a FIR filter? If not why so?
The person I was discussing this with posted the following links:
http://bigwww.epfl.ch/publications/unser9301.pdf
http://bigwww.epfl.ch/publications/unser9302.pdf
It seems as the order of the spline increases it approaches a Gaussian filter (an example of the central limit theorem at work?).
Thanks
Bill
This question came up in the new supposed big thing in audio called MQA (Master Quality Authenticated). Here is a technical overview and a link to its patent:
https://www.soundonsound.com/techniques/mqa-time-domain-accuracy-digital-audio-quality
https://patentscope.wipo.int/search...A34392194DDAEFDDC7.wapp2nA?docId=WO2014108677
They talk about explaining it with triangular sampling and linear interpolation in approximating what was lost in the sampling. This of course is an example of a spline of order 1 - pretty trivial really. But it is claimed better results are obtained using higher order splines:
'Even better results are possible using higher-order ‘B-spline’ kernels, which allow both the position and intensity to be identified of two or more separate pulses occurring within the same sampling period!'
The MQA guys are bit coy about exactly what they do, but the conjecture is they do an an analysis of the music, decide from that what order of spline to use in sampling it down to 96k and put the best way to upsample it in metadata encoded in the bit-stream in some way.
The reason is they claim to want to reduce time smear as much as possible ie when fed a Dirac impulse it has the shortest time response.
OK so far. But in the discussion I am having, some guys claim that above order 1 spline reconstruction is not a FIR (Finite Impulse Response) filter. I know polynomial reconstruction is:
https://www.dsprelated.com/freebooks/pasp/Lagrange_Interpolation.html
I know splines are polynomials between points - but not exactly the same as Lagrange interpolation. However since the whole object of this is to keep the time response as short as possible I can't see how it can be anything but FIR.
Does anyone know the answer - is it a FIR filter? If not why so?
The person I was discussing this with posted the following links:
http://bigwww.epfl.ch/publications/unser9301.pdf
http://bigwww.epfl.ch/publications/unser9302.pdf
It seems as the order of the spline increases it approaches a Gaussian filter (an example of the central limit theorem at work?).
Thanks
Bill
Last edited: