ESP32S3 deploy predictions always wrong

Question/Issue: I’ve been following the tutorial for local deployment to my ESPS3 of a keyword spotting impulse. I’ve replaced most of the functions within ESP_NN with the ones provided by espressif, otherwise the example is exactly the same as provided by Edge Impulse.

When I copy the raw features from the EI raw classifier the prediction is always wrong, usually biased towards the “unknown” label. It does not matter if I try to deploy the 8 bit integer optimized model or the float model.

Am I doing something wrong?

Project ID: 265677

Context/Use case: Keyword spotting device on the edge.

I’m having the same problem, can’t build it even when following some guides. It successfully builds if it’s set-target is a esp32 but not a esp32s3 :confused:

1 Like

@Eudald & @ZKBX

I don’t have an S3 but I would start here:

  • Check this forum post.
  • Check all links at this page.

Have you tried to replace the ESP_NN folder with one optimised for the S3, if so consider the following:

I had problems building the example for the S3 on Linux because the ESP_NN that I got had the extension of the S3 files in upper case (.S instead of .s) but the CMake file that comes with the example included those as lowercase and could not find them. I think this would work fine on Windows, but not on Linux, since Linux is case sensitive with it’s paths.

I solved it by renaming all *.S files to *.s.

Thank you, but that is not my problem. I’ve been able to build the example with hardware acceleration for the esp32s3, my problem is that when I run the example, no matter the data I feed to the impulse, an unknown is always detected.

Hi @Eudald,
take a look at esp32s3 keyword project to see how you
can use edge impulse with esp32s3.

1 Like