Skip to content

edge_impulse.md

Latest commit

 

History

History
41 lines (28 loc) · 1.14 KB

edge_impulse.md

File metadata and controls

41 lines (28 loc) · 1.14 KB

Edge Impulse Compatibility

This repository's public/default build uses a stub inference backend. Local real Edge Impulse builds are supported through prj.local.conf and a compatible deployment under third_party/edge_impulse/embr.

Required Labels

A compatible classification model must emit these exact labels:

  • embr_glow
  • embr_sleep
  • embr_flip
  • unknown
  • noise

Required Audio Parameters

The app is currently aligned to:

  • 16000 Hz audio input
  • 1000 ms inference window
  • 250 ms window increase / stride

In repo terms, this means:

  • 16000 samples per inference window
  • 4000 samples per slice
  • 4 slices per inference window

If your model uses different labels or windowing parameters, update the app code to match before using it.

Command Selection Behavior

embr selects the highest-scoring label after each inference window. unknown is the default outcome, and ties for the best score resolve to UNKNOWN.

embr_glow, embr_sleep, and embr_flip are actionable commands and are forwarded to the transport layer. unknown and noise are valid no-op outcomes; they do not produce transport messages.