<The program couldn't tell the difference between a hive and my kitchen table (that was the first thing I tried it on)>
1). It’s not trained on the absence of bee sounds, and your table presumably doesn’t have any sick bee patterns to recognize. In the absence of any colony sounds, the 8 pattern recognizable patterns will pick up some background sounds and will try to make a match. Typically, it will report some insignificant possibility for one or more factors. It is not uncommon for the app to conclude that something like a table is queenless.
2) The app is not calibrated yet. That’s the whole purpose of providing the app. It should not and will not accurately recognize sick colony sounds until we get back sufficient numbers of calibration samples. That is particularly important to our being able to see which brands and price points of phones can give us accurate recordings. Obviously a new top of the line, brand name like Samsung Galaxy or iPhone, should have better microphones and sound capture cards than older and cheap Chinese knock offs. We need recordings from all types of phones to see which will work, which will not.
3). We have discovered that the built in microphones of many phones do not have a flat response and may have the frequency band truncated to the main frequencies of the human voice. That is why we ask people who are serious about helping us to calibrate the app to use one of our recommended external microphones.
4). Testing the app, sending us the recording, analysis, and an inspection report for the hive, with a note about the Table, would have been helpful. At the very least, it would have told us what brand and model phone you were using. In the absence of feedback, we can not improve performance. Submission of test files, with inspection reports, and notes, would have been useful data.
We use the failures for training, as well as the successes. Both are critical to correct classification.
James is also right, ANNs do not learn on the phone, but at this stage, our goal is to get good quality recordings, analysis reports, and colony inspection reports. Control (Inspected and known to be ok for the 8 variables, and dead out colony recordings are as important as recordings of colonies with a problem.
Success. To date, we have enough data sets that ee
We use correct dia
***********************************************
The BEE-L mailing list is powered by L-Soft's renowned
LISTSERV(R) list management software. For more information, go to:
http://www.lsoft.com/LISTSERV-powered.html
|
