I want to build a large microphone array, say 64 microphones.   Thinking to connect 16 PlayStation Eyes with a Linux machine.   Do you think if it it possible to capture all the audio streams?

Very Good !

I was starting to research the algo for sound direction-detection and this is an already made prototype. Great !

One thing has me a bit confused though, the input is 4 channels (2x2 mics) but only 2 channels are used ? (dropping ch-2&3 and only using ch-1&4 ) or I'm mis-reading the code (?)

Also, there doesn't seem to be any frequency, filtering of the input sound waves, according to what I've read, the phase-differential of arriving waves at the mic(s) is dependent (among other things) on the frequency of the sound-wave. (perhaps that's what causes the sometimes 'erratic' results ) together with the time-difference of the sampling (i.e. channels are sampled one at the time, so channel-1 is 3 time-samples ahead of channel-4 ) may induce enough jitter to bias the calculations.

Thank you very much for collating and publishing this code !

Thanks for the kind words!

I did drop channels 2 & 3.  The microphone arrays I am using only have 2cm between microphones, so I just take the samples from microphones 1 and 4 to get 6cm of separation.  I do this so that the time delay will be greater, and I can then achieve higher angular resolution.

I need to look into filtering the input, that is an excellent idea for future work. 

It is my understanding that during sampling all four channels are sampled sequentially, as you say.  However, the time difference from one channel to the next is much smaller than the time difference from one sample to the next.  Stated another way, I think that the time difference among channels during one sample (a "frame" as alsa uses that term) and the time difference from one sample to the next are so different as to make the difference among the channels in a frame negligible. 

Let me know if I can clarify any thing more, and thanks for checking this out.  I hope it proves useful!

Ahhh, That makes sense. 2 cm is way too close.

Yes, Time difference may be negligible, (sorry, most of my readings were focused on high-frequency sampling where such delays would introduce very significant phase delays).

As per frequency filtering, the algorithm used works differently than what I've read so far, it seems to work on time-domain (cross-correlating the waveform in time,and extract the time difference between them) instead of phase/frequency domain (cross-correlate the phase of the waveform to extract said time)   given that, it would appear that pre-filtering wouldn't improve it all that much. (still going through the maths thought :-}

My project is to have 2 of these on a room, (at 90deg apart) to localize the source of a speaker.

(i.e. localize the speaker within the room)   Simple usage example, within a medium size room, I say 'turn my light' and a system would know to power on my closest light  all based on the location of the voice within the room :-)

It may be also usable as an input to a filter to remove unwanted sound (much like beam-forming antennas in most wi-fi routers, where they 'direct'  the signal via software)

Current alexa/google asst. etc. give only 1 direction filtering (8 channels, but only resolves a single direction. ) with 2 of these 90deg apart, it should be possible to get an even better 'far-field' audio filtering.

If you have published any content along those likes, pls. point me to it !

Thanks,

Carlos

Your project sounds like a great application of these techniques, and I hope they help you accomplish your goals.

I haven't published anything in this field, and I am definitely not an expert.  I do enjoy learning the mathematics behind these algorithms though! 

The GCC-PHAT as it is used here as of today is performed in the frequency-domain.  Section 2 of this 2012 paper: ftp://ftp.esat.kuleuven.be/pub/SISTA/ida/reports/12-133.pdf has a good overview.  These authors claim that the computational complexity of the algorithm and its low angular resolution for small arrays make it less desirable than a time-domain alternative.  I chose to achieve higher resolution by increasing the sample rate because I found the frequency domain GCC-PHAT algorithm to be easy to understand and translate to code, but these authors propose "quadratic interpolation" as an alternative means.  I have not yet looked into such interpolation.

If I am correctly interpreting the mathematics of their alternative time-domain approach, then they are applying the definition of cross-correlation on slide 35 of http://www-pagines.fib.upc.es/~pds/Lect02.pdf, followed by whitening one of the inputs by an adaptive linear prediction filter (to maintain the unity gain advantage of PHAT weighting), copying that filter to the other input, and finally using a quadratic interpolation to increase the resolution.

Those authors seem to have moved on from this approach, as this paper: ftp://ftp.esat.kuleuven.be/sista/ida/reports/13-199.pdf from 2013 uses steered response power.  A 2016 paper (ftp://ftp.esat.kuleuven.ac.be/pub/stadius/ida/reports/CP_16-183.pdf) uses a similar algorithm augmented with a multi-channel Wiener filter to locate footsteps.

I also found this paper: http://www.handicams-fet.eu/uploads/files/6.pdf from 2015, where an eigenvalue decomposition is used for each array.  Their procedure was to use a multi-channel Wiener filter (either by a central processor or by a distributed means called DANSE) for de-noising, followed by subspace estimation in the frequency domain, and finally wide-band direction of arrival estimation.  The authors here cite the paper: ftp://ftp.esat.kuleuven.ac.be/sista/doclo/reports/01-30.pdf as a resource for understanding multichannel Wiener filtering. 

When I began my research at the start of this project, I came across these subspace-based methods, but my inability to follow the mathematics involved in the algorithms has kept me from attempting them.  For example, I have never applied the expected value operator to matrices.  I tried to follow the explanation given here: https://math.stackexchange.com/questions/642291/expectation-operator-on-a-matrix, but I think I need a more concrete example.  I may attempt to write a toy example this week in an effort to familiarize myself with the linear algebra concepts needed by these algorithms.

I hope these resources help you, and I would appreciate any resources you know of that show any related linear algebra procedures.

P.S. I am replying to this older message because I don't see the option to reply to your most recent comment.