First of all, I've seen a similar thread, however it's a bit different to what I'm trying to achieve. I am constructing a robot which will follow the person who calls it (3D sound localization). My idea is to use 3 or 4 microphones - i.e. in the following arrangement in order to determine the from which direction the robot was called:


Where S is source, A, B and C are microphones. The idea is to calculate phase correlation of signals recorded from pairs AB, AC, BC and based on that construct a vector that will point at the source using a kind of triangulation. The system does not even have to work in real time because it will be voice activated - signals from all the microphones will be recorded simultaneously, voice will be sampled from only one microphone and if it fits the voice signature, phase correlation will be computed from the last fraction of second in order to compute the direction. I am aware that this might not work too well i.e. when the robot is called from another room or when there are multiple reflections.

This is just an idea I had, but I have never attempted anything like this and I have several questions before I construct the actual hardware that will do the job:

  1. Is this a typical way of doing this? (i.e. used in phones for noise cancellation?) What are other possible approaches?
  2. Can phase correlation be calculated between 3 sources simultaneously somehow? (i.e. in order to speed up the computation)
  3. Is 22 khz sample rate and 12bit depth sufficient for this system? I am especially concerned about the bit depth.
  4. Should the microphones be placed in separate tubes in order to improve separation?
  • 1
    $\begingroup$ Here's an interesting article, maybe you've seen it. It looks like the author ended up putting a fourth mic above the other 3 in order to deal with the sound source being above the array. Other than that it looks pretty similar to your plan (to my untrained eye, at least). $\endgroup$
    – Guest
    Feb 25, 2018 at 20:42
  • $\begingroup$ The general term for the phase correlation part is Beamforming. A common beamforming system uses a linear array of microphones, and I'm not sure the field of "vision" for your microphones will really allow for much triangulation. $\endgroup$
    – pscheidler
    Feb 26, 2018 at 4:44
  • $\begingroup$ Regarding triangulation, I guess you could set up two or three of the arrays some distance apart and find the intersection of the beams. Could solve 2-beam degenerate case with "hey robot..." (robot turns to face you)... "come here!" $\endgroup$
    – Guest
    Feb 26, 2018 at 18:30
  • $\begingroup$ Actually, that could work by adding one more mic. Check this out, it's a variation of Harry's solution. The equilateral triangle becomes a right triangle, and one more mic is added to form another triangle. From each triangle we cast a beam, and take the average of those two beams to get an accurate direction vector. Notice the two "eyes" in the demo. They're placed so that the beams running through them will triangulate position when source is directly in front of or behind robot. Try it out with source at any y=0. $\endgroup$
    – Guest
    Feb 27, 2018 at 5:36
  • 1
    $\begingroup$ @FilipePinto have you read the answers and the description of the problem thoroughly? It can't really work like that since you can't know how each energy peak from each microphone is correlated with other microphones - that's why you need phase correlation, iterative closest point or some other registration algorithm (registration doesn't refer to recording here, but to matching one signal against another) to match recorded waveforms and detect their mutual shift within some time window $\endgroup$ Jul 3, 2019 at 13:22

2 Answers 2


To extend Müller's answer,

  1. Should the microphones be placed in separate tubes in order to improve separation?
  1. No, you are trying to identify the direction of the source, adding tubes will only make the sound bounce inside the tube which is definitely not wanted.

    The best course of action would be to make them face straight up, this way they will all receive similar sound and the only thing that is unique about them are their physical placements which will directly affect the phase. A 6 kHz sine wave has a wavelength of $\frac{\text{speed of sound}}{\text{sound frequency}}=\frac{343\text{ m/s}}{6\text{ kHz}}=5.71\text{ mm}$. So if you want to uniquely identify the phases of sine waves up to 6 kHz, which are the typical frequencies for human talking, then you should space the microphones at most 5.71 mm apart. Here is one item that has a diameter that is less than 5.71 mm. Don't forget to add a low pass filter with a cut-off frequency at around 6-10 kHz.


I felt that this #2 question looked fun so I decided to try to solve it on my own.

  1. Can phase correlation be calculated between 3 sources simultaneously somehow? (i.e. in order to speed up the computation)

If you know your linear algebra, then you can imagine that you have placed the microphones in a triangle where each microphone is 4 mm away from each other making each interior angles $60°$.

So let's assume they are in this configuration:

      / \
     /   \
    /     \
   /       \
  /         \
 A - - - - - B

I will...

  • use the nomenclature $\overline{AB}$ which is a vector pointing from $A$ to $B$
  • call $A$ my origin
  • write all numbers in mm
  • use 3D math but end up with a 2D direction
  • set the vertical position of the microphones to their actual wave form. So these equations are based on a sound wave that looks something like this.
  • Calculate the cross product of these microphones based on their position and waveform, then ignore the height information from this cross product and use arctan to come up with the actual direction of the source.
  • call $a$ the output of the microphone at position $A$, call $b$ the output of the microphone at position $B$, call $c$ the output of the microphone at position $C$

So the following things are true:

  • $A=(0,0,a)$
  • $B=(4,0,b)$
  • $C=(2,\sqrt{4^2-2^2}=2\sqrt{3},c)$

This gives us:

  • $\overline{AB} = (4,0,a-b)$
  • $\overline{AC} = (2,2\sqrt{3},a-c)$

And the cross product is simply $\overline{AB}×\overline{AC}$

$$ \begin{align} \overline{AB}×\overline{AC}&= \begin{pmatrix} 4\\ 0\\ a-b\\ \end{pmatrix} × \begin{pmatrix} 2\\ 2\sqrt{3}\\ a-c\\ \end{pmatrix}\\\\ &=\begin{pmatrix} 0\cdot(a-c)-(a-b)\cdot2\sqrt{3}\\ (a-b)\cdot2-4\cdot(a-c)\\ 4\cdot2\sqrt{3}-0\cdot2\\ \end{pmatrix}\\\\ &=\begin{pmatrix} 2\sqrt{3}(b-a)\\ -2a-2b-4c\\ 8\sqrt{3}\\ \end{pmatrix} \end{align} $$

The Z information, $8\sqrt{3}$ is just junk, zero interest to us. As the input signals are changing, the cross vector will swing back and forth towards the source. So half of the time it will point straight to the source (ignoring reflections and other parasitics). And the other half of the time it will point 180 degrees away from the source.

What I'm talking about is the $\arctan(\frac{-2a-2b-4c}{2\sqrt{3}(b-a)})$ which can be simplified to $\arctan(\frac{a+b+2c}{\sqrt{3}(a-b)})$, and then turn the radians into degrees.

So what you end up with is the following equation:


But half the time the information is literally 100% wrong, so how.. should one.... make it right 100% of the time?

Well if $a$ is leading $b$, then the source can't be closer to B.

In other words, just make something simple like this:

     possible_center_direction=240; //A is closest, then B, last C
   }else if(a>c){//a>c>b
     possible_center_direction=180; //A is closest, then C last B
     possible_center_direction=120; //C is closest, then A last B
     possible_center_direction=60; //C is closest, then B, last A
   }else if(a>c){//b>a>c
     possible_center_direction=300; //B is closest, then A, last C
     possible_center_direction=0; //B is closest, then C, last A

//if the source is out of bounds, then rotate it by 180 degrees.

And perhaps you only want to react if the sound source is coming from a specific vertical angle, if people talk above the microphones => 0 phase change => do nothing. People talk horizontally next to it => some phase change => react.

$$ \begin{align} |P| &= \sqrt{P_x^2+P_y^2}\\ &= \sqrt{3(a-b)^2+(a+b+2c)^2}\\ \end{align} $$

So you might want to set that threshold to something low, like 0.1 or 0.01. I'm not entirely sure, depends on the volume and frequency and parasitics, test it yourself.

Another reason for when to use the absolute value equation is for zero crossings, there might be a slight moment for when the direction will point in the wrong direction. Though it will only be for 1% of the time, if even that. So you might want to attach a first order LP filter to the direction.

true_true_direction = true_true_direction*0.9+source_direction*0.1;

And if you want to react to a specific volume, then just sum the 3 microphones together and compare that to some trigger value. The mean value of the microphones would be their sum divided by 3, but you don't need to divide by 3 if you increase the trigger value by a factor 3.

I'm having issues with marking the code as C/C#/C++ or JS or any other, so sadly the code will be black on white, against my wishes. Oh well, good luck on your venture. Sounds fun.

Also there is a 50/50 chance that the direction will be 180 away from the source 99% of the time. I'm a master at making such mistakes. A correction for this though would be to just invert the if statements for when 180 degrees should be added.

  • $\begingroup$ I wonder if the phase thing is really necessary, or if each mic can just look for some identifiable feature. If all mics hear "hey robot" then couldn't they line up the onset of that "bah" sound and ignore phase? Then you shouldn't need to place the mics so close together... $\endgroup$
    – Guest
    Feb 25, 2018 at 22:25
  • 1
    $\begingroup$ @HarrySvensson, I see what you mean. I was thinking you could use something like your approach, except $a$, $b$ and $c$ would be a number of milliseconds since the first mic heard the sound. I played around with it here, but it's not lining up perfectly when the source, a mic, and the center of the robot aren't all in a line. I think it might be "okay" though, check it out. Error's not as bad when source is far from mics. I'm sure it could be corrected, but the math escapes me. $\endgroup$
    – Guest
    Feb 26, 2018 at 6:59
  • 1
    $\begingroup$ Not sure I've ever seen code highlighting working here on SE.DSP. Let me check with the Teacher's Lounge and see what they say. Looks like someone asked on Meta some time ago, but no action was taken: dsp.meta.stackexchange.com/questions/133/… $\endgroup$
    – Peter K.
    Feb 26, 2018 at 14:04
  • 1
    $\begingroup$ Please go and upvote that post on Meta.DSP. I've added the tag <kbd>feature-request</kbd> which should at least see some engagement, but we need the votes. If the Chemistry.SE site has it enabled, we should definitely! :-) dsp.meta.stackexchange.com/questions/133/… $\endgroup$
    – Peter K.
    Feb 26, 2018 at 14:21
  • 1
    $\begingroup$ @endolith You are correct, I have now removed that part. Thank you. $\endgroup$ Jan 28, 2020 at 3:39
  1. Yes, this feels reasonable and typical.
  2. You can just as well use the three microphone signals at once (not going the "detour" through your three pair correlations). Look for "MUSIC" and "ESPRIT" in direction-of-arrival applications.
  3. Very likely it is. You're not aiming for high audio quality, you're aiming for good corss-correlation properties, and a few bits here and there will probably not make or break the system. A higher sampling rate like the very common 44.1 kHz or 48 kHz, on the other hand, would instantly double the angular precision, very likely, on same observational length.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.