OK, that's different! But much more complicated. In this case, it takes minimum 3 microphons to determine the direction unambiguously. If I were faced with this problem, I first would build an environment to test different algorithms and only then try to solve the original task. Anyway, it is challenging. Even for a professional engineer.
I can put a first question: imagine a source of noise in a certain distance right in front of the two micros. Both signals are equal. When this source moves to a side, signal runtime (and strength) will change. The phase shift is proportional to runtime variation and frequency. If you apply FTT, both spectra should show the same amplitudes, but different phases. How to determine the frequency dependence of the phaseshift?
Is this all part of the plan to build a shooting target 'hit position detector' with three microphones ? If it is, it may be productive to say so people have an idea of the end application and a clearer view of where you're trying to go.
Transmit a sign wave and then on the input side of things look at the zero crossings, that is essentially the phase shift except you don't need the convolution. The zero crossings can be created in software after ADC or with a comparator in hardware.
It is not obvious, but convolution and deconvolution is a very basic process, and topic of nearly all research efforts. Let us have a point-shaped source of light moving in a certain distance from left to right . The brightness will not change much. Then put a slot in between and do the same experiment. What do you see? The slot covers the source, night is followed by "sunrise", "hight noon", "sunset" and night again. This is equivalent to a slot moving in between the observer and a light source. So, you also could ask: what looks the slot like? What are his dimensions? In our case, the width of the slot can be calculatet, if we now speed and distance. But what, if the light source is not a point, but is extended? And what, if there are some sources in a line?
The smaller the slot is, the better the light sources are discriminated. But also the signal strength goes down.
What we actually see is wether the lightsource nor the slot, but a signal convoluted from light source AND slot!
Now one could think, that, if the slot is known, we can deconvolute the signal and determine position and dimension of the light source. But:
convolution is an weighted integration, that is, we add different single values, being the product of two signals. And : integration creates a kind of mean value. And this is a way to reduce noise. Therefore, deconvolution creates noise and hides the signal! That is the reason, why there is no simple solution, for you don't get a dollar song for a nickel!
ErNa: Yes 3 microphones are required for full 360 but two is enough 180 degrees and after that it's just a matter of repeating the process. with 3 microphones you can even find the distance of the source.
Hippy: Yes this is a completely different person but trying to do something similar. Although I would never use it for that application and I beleive that sound is the worst choice of technology in a shooting range. And afterall doing impulses is much easier than continuous sound.
Graham: I have already tryied that (with descreetes). Problem is tha although this method works fine with sinosoids and in theory its is very prone to interaural intensity differences. What I did is use two comperators with very low ref voltage(practically 1 bit ADC) and feed them to an XOR. Then I low passed the output of the gate. The integrated output of the PWM signal should raise in maximum time defferences and be minimum when the waveforms are alike (when the source is in the middle for instance).
My protype gave bad results. And if you put the preamps and shift registers (for the delays) in the picture you can easily see it gets too bulky and power consuming for any practical use.
Thank you for taking time to share your ideas guys. Sincerely.
I was suggesting you measure the phase difference directly by timing the time between zero crossings, take averages of these to improve the result.
Personally I think convolution is a convoluted way of going about this and there must certainly be analogue circuits that will also give you a signal based on phase.
Indeed, the task is not simple, that is, can't be done (perfectly). Analog circuits bring additional problems, like noise. A digital solution will be the best, but, it will be for example, more complex then echo cancelation! So, it might make sense to look for those algorithms!
Comments
I can put a first question: imagine a source of noise in a certain distance right in front of the two micros. Both signals are equal. When this source moves to a side, signal runtime (and strength) will change. The phase shift is proportional to runtime variation and frequency. If you apply FTT, both spectra should show the same amplitudes, but different phases. How to determine the frequency dependence of the phaseshift?
No, that was someone else. Apologies.
Post Edited (hippy) : 12/5/2007 9:11:47 AM GMT
Transmit a sign wave and then on the input side of things look at the zero crossings, that is essentially the phase shift except you don't need the convolution. The zero crossings can be created in software after ADC or with a comparator in hardware.
Graham
The smaller the slot is, the better the light sources are discriminated. But also the signal strength goes down.
What we actually see is wether the lightsource nor the slot, but a signal convoluted from light source AND slot!
Now one could think, that, if the slot is known, we can deconvolute the signal and determine position and dimension of the light source. But:
convolution is an weighted integration, that is, we add different single values, being the product of two signals. And : integration creates a kind of mean value. And this is a way to reduce noise. Therefore, deconvolution creates noise and hides the signal! That is the reason, why there is no simple solution, for you don't get a dollar song for a nickel!
Hippy: Yes this is a completely different person but trying to do something similar. Although I would never use it for that application and I beleive that sound is the worst choice of technology in a shooting range. And afterall doing impulses is much easier than continuous sound.
Graham: I have already tryied that (with descreetes). Problem is tha although this method works fine with sinosoids and in theory its is very prone to interaural intensity differences. What I did is use two comperators with very low ref voltage(practically 1 bit ADC) and feed them to an XOR. Then I low passed the output of the gate. The integrated output of the PWM signal should raise in maximum time defferences and be minimum when the waveforms are alike (when the source is in the middle for instance).
My protype gave bad results. And if you put the preamps and shift registers (for the delays) in the picture you can easily see it gets too bulky and power consuming for any practical use.
Thank you for taking time to share your ideas guys. Sincerely.
Personally I think convolution is a convoluted way of going about this and there must certainly be analogue circuits that will also give you a signal based on phase.
Graham