Interesting. Am I clear in understanding that in this example, the location of the localized sound co-incidentally exactly in the center for the four recorders is ?

Very interesting screen shot, do any of the numbers in the image above represent the time of arrival of the sound event? i.e. the four numbers:

10:06.234421
10:06.577550
10:06.573095
10:07.164900

It looks like it's aligned the determination of the common sound, so then the times on the LHS would be relative to each other representing the TDOAs.

The 4x microphones appear to be in a square to the right of the Brisbane Corso Park logo in the image snippert below.

However, if I check the localization based on this information with my own software (sbts-aru, based on Tessas opensoundscape), I get the following localization:
 

$ localize_event.sh 20
Enter GPS coordinates and timestamps. Press enter twice to finish.
-27.505271,153.018879 2024-11-12_12-10-06.234421
-27.505270,153.018706 2024-11-12_12-10-06.577550
-27.505146,153.018880 2024-11-12_12-10-06.573095
-27.505145,153.018707 2024-11-12_12-10-07.164900

Location: -27.50608318833281,153.0195547475019

Web links:

OpenStreetMap: https://www.openstreetmap.org/?mlat=-27.50608318833281&mlon=153.0195547475019#map=15/-27.50608318833281/153.0195547475019

Google Maps: https://www.google.com/maps?q=-27.50608318833281,153.0195547475019&t=h&z=15



And that location is in the houses outside of the part where you did the test. 120ish meters from where the microphones are. A bit unlikely I think that the bird was heard with that spectrogram at that distance.

Later I'll look at your image and make approximations as to the center of the pattern to see if it more accurately determines where the sound is located. Unless it really is where my calculations show, but then I don't understand the picture below it that appears to show a location that is in the center of the microphones. As it appears unlikely that the location of the sound is 127m from the park.

 

However, the location co-ordinates shown in the image is indeed in the center of the microphones, but then the numbers on the left of each spectrogram can't represent the relative time of arrives of the bird sound IMHO, because the localization math based on those numbers don't return the center of the microphones in my calculations and I've tested the accuracy  of these in real life many times.

$ time_diffs.py
2024-11-12_12-10-06.234421
2024-11-12_12-10-06.577550
2024-11-12_12-10-06.573095
2024-11-12_12-10-07.164900

0.000000
0.343129
0.338674
0.930479

 However, the  difference between the earliest and the latest sound here is 0.930479s which is:

0.930479 * 343
319.154297

About 319m. Based on the co-ordinates of the microphones in the park, this is impossible, which suggests strongly that the numbers on the left of each spectrogram do not represent correctly the relative times of arrival of the sounds. The mics appear to be around 15-20m apart from each other.

Even my calculations can't be correct if the input data is correct because the location of the mics is clearly in a square around 14-20m on a side, but the max distance on the time differences is around 300m. My conclusion would be that one or more of the sounds that are shown in the spectrograms are not from the same tweet. But if that were the case, then the example in the botton with find on map could not possibly resolve to the center of the microphones as it appears to??

The location co-ordinates shows in the botton frame on the find on map image clearly resolves to the center of the microphones as shown in the image. How can there then be 0.9s difference in the times ?? I think I must be missing something.

Michael, are you able to confirm whether something in the images is not correct? Such as they are not the same tweet ? Or that I'm miss-interpreting the time numbers on the left of each image?

Or.... the time numbers on the left that we are seeing are the ones that do not compensate for the time drift or something like that.

OK, only @MichealMaggs knows, but my interpretation of the figure is that the times given left are the uncorrected times and the after the correction the arrival times are more or less equal, (zooming into the figure shows that they are not exactly aligned) and consequently the source location is in the middle of the microphone array. Also the top figure indicated that is for selecting the sound event for correlation and not for visualizing the correlation result.

Ah okay that make's sense. Yes, start time of the whole file would of course have different drift between them all. Though if the file sizes are something like 10 minutes, 0.9s would be a lot of drift for that short time. Indeed, I agree, the times in the middle are very close, suggesting that it is connected to a localization in the center of the mics, plus the spectrogram shows the volume is quite high, confirming that. Maybe the file lengths are much longer than 10 minutes?

Yes, the top image is for aligning the events, but it does look like after aligning the times on the left represent the relative TDOAs though, which allowed me to make a localization on the assumption that they were corrected values.

Indeed, the difference in time on the middle values is 0.004455, which translates to about 1.5m distance, so they are aligned well.

It could also be a loss of satellite lock on two of recorders for a while, but if that were case then the math would not reveal a location at the center of the mics as it has. So this would suggest that your theory about the times that are visible being the uncorrected values is the most likely.

Do you know if the correlation process is fully automatic or manually assisted? It's done a very nice job of aligning them I think.

Thanks for your insights!

Hi Kim,

I love that you fact checked it! 

Yes the call I showed is from a ground truthing test I did in the park at Fairfield. (I couldn't be bothered censoring out the location info in more interesting dataset to show but I think I'll go do that now...). In the test I made sounds at different locations inside and outside the array to measure the accuracy and see how the performance changed relative to the recorder positions. The short conclusion is that everything inside the array was pretty good and relatively similar and everything outside the array looked more like shadows around a campfire. Basically what you’d expect from beamforming kinda stuff.

The four numbers in the green boxes aren’t the timestamps of the calls, they just show the file-time at the mouse cursor (vertical green line). All of the audio is shown time aligned and stretched by a tiny amount (to compensate for crystal drift), so the cursor just shows where you’d go looking in the actual audio file for whatever in under the cursor. The “real time” at the cursor is shown in the bottom left. (It just auto-hides after a few seconds when there is no cursor movement, like in those screen shots).

If you want the TDOA’s to plug into your package I can get those too. See below. 

I tend not to use those as it depends very heavily on how well your cross-correlation (or whatever method) is able to work out the time offsets, plus if your recorders aren’t well synchronized then it’s hard to tell if you have a good result or not since a single lat long output has lost a lot of juicy, noisy information. 

For example, below is the same sound localized with accurate time synchronization (Left) and with artificial time errors introduced (Right). From the outside they both look like they could be correct, but you have no way to really know. 

To overcome this I prefer to use a “cross-correlation map” because errors in time or ambiguities in the signal that would introduce errors into the result are easy to spot in the output. See below.

This is the same sound as above with accurate time synchronization (Left) and with artificial time errors introduced (Right). On the left you can see the lines converging on a single point whereas on the right you can see that they do not converge. 

Furthermore, some sounds are harder to localize because the signal is too noisy. Using TDOA you can’t tell if the result is “good” or accurate but with the cross-correlation map you can spot a bad signal.

Below is a comparison of the two methods for the same noisy signal. 

TDOA is use on the left and CEM on the right. TDOA gives you a false sense of accuracy whereas CEM is easier to see that the result is noisy and not to be trusted. 

I hope that wasn't too long winded of a response. As you can probably tell I could nerd out about this stuff all day.

As for some more exciting results here are some censored pics I can share:

This one is of a Koala. I also have an animated version of this that shows a bat (Austronomus australis) flying in and doing some feeding passes before leaving.

 

And these last two are of Australasian Bitterns.

This is a Bittern calibration test designed by Eirene Carajias and Karen Rowe where we went around in this area and played multiples of different bittern calls at different locations to see how well the replicates clustered. Each coloured pin is a cluster of results staked on top of each other. This is just a small subset of the results. Eirene did a mammoth job and filled the who field with data points! 

Absolutely not too much detail @MichaelMaggs  ! Those cross correlation maps look really interesting, well done. I'll have to have a play with those myself. I had never heard of them before, maybe opensoundscape can do some of this stuff.

Thanks for taking the time to provide this details :-)

Assuming that you had the source of the sounds is within the polygon of microphones, at what kind of distance are you able to sound localize say Koalas well ?

Hi @kimhendrikse,

With our recorders (mic quality is important!), and depending on terrain, their calls can easily be seen on a spectrogram 400 - 700m away, which has been determined with data from large arrays and recognizable individual calls appearing on multiple recorders.

So if your recorders were spaced ~300m away from each other (an offset grid of equilateral triangles is the best for overlay efficiency but certainly not always practical) then you should be ok. But I've not confirmed this distance experimentally with triangulation yet. I need to touch base with Richard Hopkinson and Roger Martin again as they are currently trialing this on Koalas. We're going to run our own Koala trials soon hopefully and see how far we can push recorder spacing.

Okay, super! Really nice thread you started. It’s great to get these insights.

With my own experiments, as you mentioned, accuracy is very good within the polygons. But in addition I found surprisingly good distance outside of the polygons in the case of a square layout along the axis of the square beyond the vertice.

I concur about the quality of the microphones. I’m using microphones that get good reviews from people who do nature recordings.

Looking forward to hearing about the Koala trials.