so is the TAS coefficient precision only 23 bits or not ?
here is the analysis of the sharcs capability Extended Precision Floating vs Fixed Point - Sharc | Comp.DSP | DSPRelated.com
TI's documentation is clear about the hardware capabilities. You'll have to ask Hypex as to the scaling details of their implementation.
For the money I would have expected something more like FIR and a World Clock Input. XMOS USB chip or similar. What do I get for the much higher price compared to the Mini DSP more?
Hi,
Don't understand little things (don't undestand anythings about digital 😱) and have please just two questions:
if the DLCP is a preamplifier too, where is the volume control ? On the software...on the computer which must be switched on all the time) ? (if a cd player or an individual steramer is connected via spdif). in a case of a streamer like a squeezebox duet used with spdif output: can I use the volume controle of the streamer without any loss ?
Understand that Hypex is not full "digital" like the little FDA amp of HifiMeDiy DDX320 with its little remote ant its 14 tone curves (of course I don't want here to compare these two products which are not made for the same people). But if the DLCP is made in relation to the Ncore amp first and for the others brand after, is there possibility to make shorter between the DLCPs' DACs--> OPA1632 and the "analogic" input stage of the hypex amp and its amp stage ?
Sorry for the low technic level questions, not sure to understand the different topologies about digital amplifier...
Don't understand little things (don't undestand anythings about digital 😱) and have please just two questions:
if the DLCP is a preamplifier too, where is the volume control ? On the software...on the computer which must be switched on all the time) ? (if a cd player or an individual steramer is connected via spdif). in a case of a streamer like a squeezebox duet used with spdif output: can I use the volume controle of the streamer without any loss ?
Understand that Hypex is not full "digital" like the little FDA amp of HifiMeDiy DDX320 with its little remote ant its 14 tone curves (of course I don't want here to compare these two products which are not made for the same people). But if the DLCP is made in relation to the Ncore amp first and for the others brand after, is there possibility to make shorter between the DLCPs' DACs--> OPA1632 and the "analogic" input stage of the hypex amp and its amp stage ?
Sorry for the low technic level questions, not sure to understand the different topologies about digital amplifier...
Last edited:
Some random clarifications.
Coefficient length is rarely an issue. Coefficient rounding only causes minor errors in the frequency response which are only noticeable when you make very high Q's at low frequencies. 28 bit coefficients are more than enough for any use sensibly related to loudspeakers. What really matters sonically is the length of the data words. Data rounding translates directly as distortion. In IIR filters you very often find yourself subtracting two nearly identical numbers and then amplifying the rest. Single precision floating point is slightly better than 24 bits fixed because at the very least the error scales with the signal level. But then again for most practical purposes 32-bit fixed is always better than single precision float - float only wins when the signal level drops below -48dB.
The data word lenth in the TAS3108 is 56 bits (48 after the point) and its accumulator is even longer. That's the main reason why I love it so. It's also got a cute microcoded architecture. You can seriously geek out writing code for this chip. I'm seriously galled that the 3108 has been relegated to "not recommended for new design" with no alternative in sight. The "alternative" given on the TI site is ridiculous. What this means is that we'll be doing a lifetime buy when they announce end of production.
The SigmaDSP chips have a similar microcoded architecture but its word length is limited to 28 bits (data and coeff), with limited support for double precision. If you want to do double precision you have to manipulate half words yourself which doubles the number of cycles needed. They're no alternative for the 3108 but their built-in converters and more flexible I/O give them their proper niche.
For those still believing that class D should be closer to digital than a linear amplifier, please read http://www.hypex.nl/docs/papers/AES120_353BP.pdf. The notion of "digital amplifiers" is nonsensical. The only thing you can conceivably eliminate between a DAC and an Ncore amp is the input buffer stage if you change the gain of the DAC's post filter. That would mean a different filter for each prospective DLCP/amp pairing, which is not economical.
As amply attested, I only believe in FIR stages as a way to approximate poles outside the unit circle targeted specifically to cancel non-minimum phase zeros that inevitably result from summing higher order LPF/HPF slopes. There simply aren't justifiable uses of FIR otherwise. Some future update of the software and firmware will support FIR approximation of noncausal IIR. You don't need a lot of DSP for that. The LS1 has a 6000 tap filter correcting the 70Hz subwoofer crossover, and it's done in a hundred cycles or so. The main question will be how to make the software intelligently trade delay memory for this use and for time alignment. We know how to do it, just not whose time to do it in.
Coefficient length is rarely an issue. Coefficient rounding only causes minor errors in the frequency response which are only noticeable when you make very high Q's at low frequencies. 28 bit coefficients are more than enough for any use sensibly related to loudspeakers. What really matters sonically is the length of the data words. Data rounding translates directly as distortion. In IIR filters you very often find yourself subtracting two nearly identical numbers and then amplifying the rest. Single precision floating point is slightly better than 24 bits fixed because at the very least the error scales with the signal level. But then again for most practical purposes 32-bit fixed is always better than single precision float - float only wins when the signal level drops below -48dB.
The data word lenth in the TAS3108 is 56 bits (48 after the point) and its accumulator is even longer. That's the main reason why I love it so. It's also got a cute microcoded architecture. You can seriously geek out writing code for this chip. I'm seriously galled that the 3108 has been relegated to "not recommended for new design" with no alternative in sight. The "alternative" given on the TI site is ridiculous. What this means is that we'll be doing a lifetime buy when they announce end of production.
The SigmaDSP chips have a similar microcoded architecture but its word length is limited to 28 bits (data and coeff), with limited support for double precision. If you want to do double precision you have to manipulate half words yourself which doubles the number of cycles needed. They're no alternative for the 3108 but their built-in converters and more flexible I/O give them their proper niche.
For those still believing that class D should be closer to digital than a linear amplifier, please read http://www.hypex.nl/docs/papers/AES120_353BP.pdf. The notion of "digital amplifiers" is nonsensical. The only thing you can conceivably eliminate between a DAC and an Ncore amp is the input buffer stage if you change the gain of the DAC's post filter. That would mean a different filter for each prospective DLCP/amp pairing, which is not economical.
As amply attested, I only believe in FIR stages as a way to approximate poles outside the unit circle targeted specifically to cancel non-minimum phase zeros that inevitably result from summing higher order LPF/HPF slopes. There simply aren't justifiable uses of FIR otherwise. Some future update of the software and firmware will support FIR approximation of noncausal IIR. You don't need a lot of DSP for that. The LS1 has a 6000 tap filter correcting the 70Hz subwoofer crossover, and it's done in a hundred cycles or so. The main question will be how to make the software intelligently trade delay memory for this use and for time alignment. We know how to do it, just not whose time to do it in.
Back a while ago you stated that linear phase was overrated and that the advantage of using the FIR correctional filter in the LS1 didn't affect sound much at all. You theorized though that it would be more audible at lower frequencies.
So, now that you've designed the LS1s, did you find that this prediction was correct? Was FIR correction was more of an advantage lower down or is it hard to detect the improvement there also?
// Olle
I've not yet AB'd. I've recently added a function to turn the phase correction on and off at will though so I should be able to do so easily now.
You should get much more audible results with the phase linearization of this 70hz crossover than the ones you got with the 1.5khz crossover, according to the majority of reports I have from users of rePhase.
(by the way Bruno, this was your LS1 white paper that gave me the motivation to write this software 🙂 )
(by the way Bruno, this was your LS1 white paper that gave me the motivation to write this software 🙂 )
Hi Bruno
First of all I'd like to wish you a Happy New Year.
Second:
One transient-perfect opology that wouldn't be too hard to implement is the subtractive-delay one. That would be much easier for the users and the programmer compared to FIR or other pseudo-noncausal methods.
Regards
Charles
First of all I'd like to wish you a Happy New Year.
Second:
One transient-perfect opology that wouldn't be too hard to implement is the subtractive-delay one. That would be much easier for the users and the programmer compared to FIR or other pseudo-noncausal methods.
Regards
Charles
Hi Charles, vir jou dieselfde!
So far the subtractive topologies I've seen all trade in phase matching for improved transient behaviour on-axis. I wouldn't know what Psi are doing though, they seem to have figured it out quite well.
So far the subtractive topologies I've seen all trade in phase matching for improved transient behaviour on-axis. I wouldn't know what Psi are doing though, they seem to have figured it out quite well.
Dankie !
PSI is using conventional crossovers and a classic phase-equaliser (built from 2nd order allpass sections) in front of them.
What I meant by subtractive-delay is something like that (second picture of 2nd post):
Problem with subtractor block | EngineerZone
If you use a higher-order Bessel for the lowpass you will end up with quite good phase-matching between the different ways. But the Highpass wouldn't be very steep.
Regards
Charles
PSI is using conventional crossovers and a classic phase-equaliser (built from 2nd order allpass sections) in front of them.
What I meant by subtractive-delay is something like that (second picture of 2nd post):
Problem with subtractor block | EngineerZone
If you use a higher-order Bessel for the lowpass you will end up with quite good phase-matching between the different ways. But the Highpass wouldn't be very steep.
Regards
Charles
Could the same responses not be approximated using the filter/delay functionality in the existing Hypex software?
There is just a tiny little bit that is missing: You would need to subtract an IIR lowpass function from a delayed (but otherwise unaltered) version of the input signal. The delay must be set to the same value as the group-delay of the lowpass at zero Hz.
Filtering and delaying alone wouldn't work.
Regards
Charles
Filtering and delaying alone wouldn't work.
Regards
Charles
OK, sorry ...
One way to correct such things would be the use of a chain of allpasses but they wouldn't be easy to dimension. I for instance do this with an Excel sheet and some trial and error. But there are of course filter design programs that can do phase-linearisation but they are quite costly, especially if you don't need them too often.
Regards
Charles
One way to correct such things would be the use of a chain of allpasses but they wouldn't be easy to dimension. I for instance do this with an Excel sheet and some trial and error. But there are of course filter design programs that can do phase-linearisation but they are quite costly, especially if you don't need them too often.
Regards
Charles
This paper here was presented at an AES meeting and shows a method that is a good compromise regarding transient accuracy and frequency response. Unfortunately it also suffers from some lobing issues but less so than the classic constant voltage subtractive crossovers:
http://www.harschacoustic.ch/Crossovers with quasi linear phase response.pdf
Regards
Charles
http://www.harschacoustic.ch/Crossovers with quasi linear phase response.pdf
Regards
Charles
I also think linear phase is overrated....
We have to see everything in it's perspective. Only a loudspeaker with 1 driver hase a linear phase, if you ad a driver it aint anymore. Phasing will be dependent of listening position and can't be corrected with a filter ( or you should put a sensor on your head attached with your dsp😛) The only thing you can do is correcting the phase on one position ( aka the sweet spot) and is done with delay. If you change your listening position in the vertical plane, the delay time for the drivers to reach your ears change and so will the phase..... The most important thing in filtering is to get the phase right at the crossover point related to the intended listening position.
The attached picture is from the earlier mentioned linear phase xover, but demonstrates this phase "problem" clearly ( and does not solve it!).
We have to see everything in it's perspective. Only a loudspeaker with 1 driver hase a linear phase, if you ad a driver it aint anymore. Phasing will be dependent of listening position and can't be corrected with a filter ( or you should put a sensor on your head attached with your dsp😛) The only thing you can do is correcting the phase on one position ( aka the sweet spot) and is done with delay. If you change your listening position in the vertical plane, the delay time for the drivers to reach your ears change and so will the phase..... The most important thing in filtering is to get the phase right at the crossover point related to the intended listening position.
The attached picture is from the earlier mentioned linear phase xover, but demonstrates this phase "problem" clearly ( and does not solve it!).
