Mobile phone classification error

Report bugs
#1

Android Mobile phone Deploy error
Android Mobile phone Deploy error1080×1920 156 KB

Error when deploying on android mobile device cough detection.

#2

Hi @AATOS,

Could you share your project ID so I can have a look a the mfcc configuration? The ID is displayed on the dashboard page.

Thanks,
Aurelien

#3

Thanks aurel for looking into my mobile error.
I also have 0 computer time use left.not sure if it is related or not.Did get also a bunch of unknown projects loaded on my project page which could have affected my computer time use.
my project ID is:
Project ID 5305

#4

Thanks for sharing.
A couple of things I noticed:

  • In your test set, you have some labels named ‘testing’, it would be good to rename them with their correct labels (cough, noise…)
  • When testing the model, there are warnings on the features’ size, most likely some samples have different sampling frequencies in your test set. Can you check the audio configuration on your devices? e.g 16 bit/16kHz mono channel. It’s important to have same audio parameters both when training and running live classification.

Thanks,
Aurelien

#5

Thanks for the tips as I was just following instruction on web link:


I think all of the testing audio files were downloaded from the weblink.
Do I then really need them.
When adding own recording I was using the default mic setting on Android phone and will see if I can set it as specify when recording audio.
thanks

#6

Hi @AATOS, at some point you’ve had mixed accelerometer and audio data in this project and the MFCC block had the accelerometer data also configured as an input. But your signal only has audio data, thus failing. You see the same thing on the DSP screen:

image
image1484×576 26.5 KB

Usually you can remove the axes from the ‘Create impulse’ screen but if there’s no accelerometer data in the project anymore those axes disappear. As a quick fix:

  1. Go to Create impulse
  2. Remove the MFCC block
  3. Add the MFCC block again

:white_check_mark:

Note that you probably want to remove your NN Classifier block and re-add it as well, so the right reference architecture is loaded. Right now the network is tuned for accelerometer data, and won’t work well for high dimensional audio data.

#7

T

Thanks @aurel and @janjongboom with your guidance, mobile cough detection is working.

coughdetection
coughdetection1080×1920 107 KB

Looking forward for next tutorials for better understanding AI applications.
EdgeImpulse is a game changer for accessible AI.
Keep up the good work.