I’ve been trying to generate a model that will work with microwakeword on esphome. all the models i’ve trained seem to have a tensor input like [1,3920] but i need a 3d input like [1,40,98] but i’m running into a brick wall everytime i run the model it seems to flatten it. unsure if it’s possible to make a 3d tensor input
Welcome to the forum @Arngrim
esphome is something I’m keen to learn more on and explore, they have a solid framework! Please share any example code you have, it would be cool to see some examples working.
Regarding your issue, can you please share the project name / id or a screenshot of your Create Impulse screen?
You may be missing a processing block, or using RAW / flatten by mistake instead of an audio one like audio MFCC or MFE that will help preserve your input.
My guess is you need to add the spectrogram block [height, width, channels] if you add this block you should be good
Best
Eoin