Servo-feedback Subwoofer

[dswiston]

TMS320C549 is in the process of being aquired, there may be a bit of work involved with getting it set up as I do not know what exactly the DSP will come with (DAC, ADC, connectors etc). Algorithm/modeling is still on the drawing board. MATLAB will be a friend of the project for the next few weeks.

[rythmikaudio]

Originally posted by dswiston

They also seem to poo-poo on the accelerometer based approach. The "mechanical" ground issue seems to be a valid one and would dictate the sub inclosure be heavy and dead to vibrations.

This mechanical ground issue goes beyond what it appears. First it is the only reason why those accelerometer-based subs can never go to cars (there is a workaround but it is not reliable because it needs 2 accelerometers with exactly same characteristic over time and temperature: one on enclosure and one on cone). I do not make this up. I had an ULD15 and each time I moved the sub when it was still powered on, the cone just went warble in and out. It took about 3 seconds to be back to normal. Anyone with other accelerometer-based sub can try the same thing. The sensing coil does not have this problem because it is referenced to the magnetic field which moves with enclosure. Second, what this induced transient response shows is that any accelerometer based sub will have a low frequency resonance below 1hz. There is a theroretical explanation for that. The phase shift from the acceleromter in an open loop sub system is 180 degrees phase shift at DC and it is the best scenario that there is no other mechanical problems to cause additional phase shift at low frequency. There is a phase compensation needs to be done. We are all familiar with compensation at high frequency. It takes time to think about how to do it for low frequency because it is not in the textbook. The movement of enclosure acts just like injecting an external stimulus and make the transient response of the resonance manifest itself. There are other source of events can also induce the transient response, such as amp clipping, power on....

As a result, while they may produce low distortion on paper, it is not necessarily reflected into the sound qualities that audiophiles are looking for.
This statement doesn't convince me of anything other than to implant a seed of doubt, after all, they are selling something.

Simplicity is the key to good sound. There are a lot of amps with ultra low distortion (with ultra high open gain and heavy feedback). But how do they compare to single-ended amps with tons of distortion? I am not saying distortion is good. However, there is a lack of understanding that types of distortions are less-pleasing to the ears. Now back to the accelerometers. Each transducer has its own sound charateristic (because of the material and construction). But no one seems to investigate that and blindly assume all accelerometers sound equally good. The piezo material is definitely not in the main stream of high end audio. We, as engineers, need simplified models of the subjects to make our jobs easier. But taking the distortion number as the only measurement of merit is going a bit too far.

An accelerometer is on order and my lab partner is working on securing a DSP that we can work with. Hopefully solid progress can be reported soon.

I know this may sound like a joke, but keep an ear muffle handy. You will need it. Most likely the first time it is power-on, it will turn into loud audible oscillation.

Originally posted by phase_accurate

Unfortunately the "mechanical ground issue" is valid for ALL other approaches as well ! So claiming that their's isn't susceptible to it is pure marketing BS IMHO.

Please see the above comment.

But the figures they present don't look that promising at all: A reduction form 15% k2 to 10% is only an improvement form very bad to bad and doesn't make one think that the principle is very effective. I know that distortion audibility is not very high at 20 Hz. But the driver excursion was only 1/4" p-p, which is still in the linear range of many mid-woofers.

First, it is seriously misquoted and taken out of context. Just because we published data with 1/4" p-p and it implies the driver has only 1/4" linear p-p excursion? The linear excursion is more than 1" p-p.

Second, we are not bothered by the 2nd order distortion. It is the 3rd and higher order that I am interested (because the audible curves of human hearing). The 2nd order distortion is simply less audible. More important, it is the spider distortion and temperature memory effect that we are trying to reduce.

Originally posted by Francis_Vaughan

Indeed, I thought the table comparing DSP with analog was either simply naive, or just plain bogus. There is nothing the DSP can't do the op-amp based design can, and the DSP has some intrinsic advantages, in particular it can include elements of time.

While DSP has a lot of advantages, it is not an all-mighty that can solve everything because of its latency and dynamic range. For instance, Can DSP replace the closing loop feedback of op amps (so that there is no need for feedback resistors)?

My view on this issue is the DSP is best suited for feedforwad error correction where latency is not an issue. And feedforward correction needs correct modeling of the subject (which is the driver) and constantly refine the models to compensate for time-to-time and unit-to-unit variations. This is possible for consistent, tight tolerance components. But speakers are not. In addition, if we look at those Tripath amplifiers, they still need closed loop feedback even though they have tons of feedforward distortion reduction. The last few miles will be from closed loop, only.

Let us keep in mind Dr. Klippel's feedforward error correction papers were published almost 8 years ago (correct me if I am wrong) and I yet to see a feasible, robust, and mass-reproduceable implementation. The difficulty is definitely there and your accessment of the number of parameters are definitely true. And personably I think the issue is inherited and cannot be resolved.

Brian D.

Rythmik Audio

[rythmikaudio]

Another point about all mighty DSP. There is another long-forgotten name for the feedforward error correction (distortion reduction) circuit (algorithms?) used by DSP --- it is called "pre-distortion" circuit. This name will give you a better idea what DSP cannot do and what closed loop feedback can do. A closed loop does not care the amount of distorion in the system. If the system does not have any distortion (or lower distortion), it works even better. Same thing cannot be say about feedforard system.

Cheers

Brian

[jcx]

i believe the beolab 5 claims voice coil resistance compensation is done by modeling the temp rise by the dsp - an example of a consumer loudspeaker feedforward correction application

[rythmikaudio]

Originally posted by jcx
i believe the beolab 5 claims voice coil resistance compensation is done by modeling the temp rise by the dsp - an example of a consumer loudspeaker feedforward correction application

Not sure if they mention how accurate their temperature measurement is, how soon the temperature is updated and how do they verify they have completely tracked the temperature? Is that based on simulation, or it is based on designed demo? More importantly, if they had known there is a closed loop feedback system that completely solves this problem without DSP, would they still use DSP?

The idea of feedforward system is that it anticipates a distortion (or characterisitic change) and inject the distortion (that is, pre-distortion) in hope it will exactly cancel the actual distortion. So there are 2 distortion components in the system, one from distortion of the system itself and one from the pre-distortion. What if they don't match because incorrect or incompletely modelling or incorrect measurement of the parameters? In this case, will hear one "natural" distortion plus one "synthetic distortion". Even worse, what if the pre-distortion component is always one step behind the actual distortion. It is like hearing some form of echo. It is true feedforward system will not oscillate. But it has its own set of problems. Servo is still managable using closed loop feedback. It is just that most people do not understand the control theory behind it, what makes it work and what makes it break. The effect of creep and flux modulation can be addressed with a closed loop feedback system.


Cheers

Brian

[dswiston]

I thought I should provide updates in case anyone is interested.

The DSP has been setup and a few function generators gathered for testing. The current scheme is using a derivative of the LMS (Least Mean Squared) algorithm, the same type used by many noise canceling devices and motor control circuits. After matlab simulations everything looked well. Using two function generators, one for the desired signal and one for the "distorted" signal the system was able to match the two signals very well. While wrentching on the voltage of one signal, the two remained close to eachother. It is not a very stressful and true to life situation though so now the sub and accelerometer have been brought into the equation while using a function generator as the input. The system proved to be unstable at first but tweaking the "mu" for the algorithm has helped and can now be run for extended periods of time. Its performance isn't where we would like it yet, but there are places for improvement. The accelerometer is only taped onto the sub and this poor coupling appears to be causing problems as it mechanically vibrates on its own. The next step is to epoxy it to the dust cap (which happens to be extremely rigid).

Further improving and at the suggestion of a professor who is very familiar with adaptive algorithms, tonight we moved from the dlms (delayed lms) to the ndlms (normallized delayed lms) which has a dynamic "mu" and converges faster than the dlms with a slight increase in computational complexity. Additionally, we are implementing a leakage factor to take into accound the fixed point precision of the DSP. Round off errors accumulate over time and "bleeding" the filter coefficients very slowly keeps it under control and helps make sure this does not cause an overflow in the filter taps.



So as you can see we haven't gone the way of a feed forward system. I wish I had time to investigate all of this but the way the class and my undergrad schedule goes, my lab partner and I had to pick a method and pretty much go with it. Making an attempt at a few different situations would end up giving me a brief experience of them all and possibly running out of time to make an at least decent functional system. Hopefully in the next five weeks we can continue improving and make the system work very well.


Random Specs:

Filter Implementation: FIR
Filter Length: 128-768 still experimenting
DSP: TMS320C549, 100MHz
Current Algorithm: ndlms with tap leakage
Sampling Rate: 44.1kHz wish I could lower this but that appears to be all that can be gotten from the board we have unless a crystal replacement is attempted
Buffer size: 512-4096 samples, still experimenting with this as well


There are quite a few algorithm parameters we are considering as well but those would prove to be too numerous to provide as we still do not have anywhere close to a "solid" range of what is going to be settled on.