In my project, I’m trying to use a microphone to determine the flow rate of air past my microphone in the presence of background noise. My input is audio data, and my output should be a numerical flow rate. My current impulse design is to split the audio into 100 ms chunks, run a MFE, and then perform regression. I am able to get fairly good accuracy with some of the models I’ve tried. It takes a while to manually iterate my DSP and NN design, so I wanted to try out EON Tuner.
Unfortunately, when I collected my training data, I did it programmatically by adding some random noise in the input flow rates. I had figured that since I’m doing regression, the more flow values I included in my input data, the better. That means that many of my labels only have 1 sample associated with them, and EON Tuner will throw an error message telling me that.
Each of my audio samples is 5 - 30 seconds long, so each one has plenty of 100 ms chunks inside of it. I could split each of those audio samples up into 5 parts and reupload them to work within the current EON Tuner constraints, but that seems a little silly.
For my use case, it would be useful if I could run EON Tuner without having to go in and modify all my samples - but I understand if that’s a workflow you might not want to support.