Guys, I’m new to machine learning and I’m starting to play around with the Arduino nano 33 ble and I got great results with words like yes, no in the samples it’s really fantastic! I would like a tip, I’m putting samples of engine noise, for example, drone, I made several samples of 1 second and 10 seconds and I couldn’t go beyond 59% it’s easily confused with noise, do you have any tips to improve? I researched in the forum and I confess that I was in doubt, if I would have to master c++ or if there would be some function that could help. thanks!
Hi @levinlinux,
Could you share your project ID? I’d be happy to take a look at your project and data setup. I suspect you probably need more data and to use a different preprocessor block (e.g. MFE instead of MFCC).
Hello friend, yes, the ID is 245338, I’ve done several tests and I confess that I managed to get good results, this weekend, I’m going to use MFE to compare! In some moments the noise is low in the samples, I suspect that it could be contaminating the training, do you have any suggestions? I thought the Edge was simply fantastic!
Hi @levinlinux,
Where are you seeing 59%? When I run “model testing” in your project, I see an accuracy of 87%. That’s a good start, but it looks like there’s some overfitting during training (comparing 87% accuracy in testing vs. 96% in training). One of the best possible fixes for this is to collect more data. I also recommend trying to get a balanced dataset to avoid having your classifier be biased toward your drone samples.
Noise is a common problem in audio classification projects. “More data” also means collecting data in your target environment. If you record your drone indoors but expect to classify drone sounds outdoors, you can expect your classifier to not work as well (as it was not trained on samples with noise from the outside).
Something you might want to consider is using data augmentation to generate new samples that mix your target sounds with various background noises. You can read more about it and see my example of mixing/data augmentation here.
1 Like
Hello Shawn, I commented above that I got good results, because I’ve trained a few times for testing, without a doubt the samples need to be improved, I’ve done several tests and I’m learning. I will study your link!
I immediately thank you very much and without a doubt this edge platform is fantastic!
Thank you very much! I’m watching the forum, there’s a lot of good information!
1 Like