Inaccurate predictions of audio model when downloaded locally on a ESP32 S3

Help
, ,
#1

Question/Issue:
[I trained an edge impulse keyword detection model. I downloaded it as an Arduino IDE library and added it to the IDE via .zip. But when I ran the static buffer example, it gave awful predictions, never the same prediction as the one from live prediction.[

Project ID:
[891372]

Context/Use case:
[The model is supposed to identify keywords from an audio captured, forward, backward, right, left, stop, noise, and unknown. I used the static buffer example to run it, but it is just not what’s printed and it is awfully wrong.

1 Like
#2

Hi @random-person1232 ,
As you might guess by the name “static buffer” example runs inference on static data. It is NOT running the inference on the actual data. Static buffer example is meant for sanity check (i.e. can you run the inference? is the prediction the same as in Studio?), not for actual usage.
For ESP32-S3, the easiest way to run inference with the microphone probably will be the Arduino sketch examples.

#3

Hi! I’m new here. I’m trying to figure out how to link trained models with hardware. Have you successfully run the model and gotten it to work with the microphone? Also, which ESP32-S3 board are you using?

#4

@fijay23 welcome to the forum!
As I mentioned above, for ESP32-S3, the easiest way to run inference with the microphone probably will be the Arduino sketch examples.
If something is not clear in that message, let me know.