I have a snoring detection project (Martin/Snoring_v2) that partially works. Its dataset has 2 classes, snoring and non-snoring. If I select both in the Posterior Parameters it works, it detects snoring sounds, and otherwise may or may not report non-snoring sounds. BTW this is using a Nicla Voice board. But if I select only Snoring in the Posterior Parameters it only reports non-snoring, usually with snoring sounds and sometimes otherwise, but never reports Snoring. Am I misinterpreting what the Posterior Parameters is asking for?
Please share the doc you are following I’m not sure about the posterier parameters. thanks!
The example I used is from hackster.io (Snoring Detection with Syntiant Neural Decision Processor - Hackster.io). The Posterior Parameter selection is in the Deployment block, selecting the Arduino Nicla Voice board and hitting the Find Posterior Parameters button (picture uploaded to
Ah ok I get you now, you usage seems to be different to the example in the guide, as using only “Snoring” in those posterior parameters is going to set it to only classify if the model is almost positive its snoring (like overfitting to snoring):
- If the model hears a sound and isn’t totally convinced it’s snoring (meaning, it doesn’t hit the confidence level you’ve set for “Snoring”), it’s going to use “Non-Snoring” by default. Not because it knows it’s not snoring, but because it’s not sure enough it is detecting Snoring.
- There’s a chance it might never set “Snoring” because you’ve asked for really high confidence to call something snoring. So, if a sound isn’t an exact match for one of the snoring sounds you used for training, it might not ever match.
Including “Snoring” and “Non-Snoring” for the posterior parameters, is giving your model a fuller picture. This way, it can weigh its confidence against the thresholds for both classes and make a more rounded call.
A little hard to picture since this is after the training, unless picking a single sets a higher confidence level than picking both. And if only snoring is picked should it even consider non-snoring vs saying nothing at all (which is often does when both are picked)?