How klumw/wakeword works
This is an ESP32-S3 wakeword-detection starter: microphone audio is captured and gain-adjusted, then passed through Edge Impulse audio processing and an embedded model; a positive “computer” detection drives the onboard LED. The README documents I2S input, software AGC, ESP-NN optimization, model replacement, and visual feedback. The sampled sources establish the classifier entry point, model artifact, DSP and inference-runtime layers; detailed capture and detection calls are not shown, so those documented links are marked unverified.
Main entry points: classifier Classifier API · keyword.cpp Keyword app · micro TFLite Micro · main Audio AGC
Components 16
Device I/O
- I2S microphone
- Keyword app
main/keyword.cpp - Audio AGC
main - Status LED
main
Audio inference
- Classifier API
components/edge_impulse/edge-impulse-sdk/classifier - Audio features
components/edge_impulse/edge-impulse-sdk/dsp/speechpy/feature.hpp - DSP math
components/edge_impulse/edge-impulse-sdk/dsp/numpy.hpp - Model metadata
components/edge_impulse/model-parameters/model_metadata.h - Compiled model
components/edge_impulse/tflite-model/tflite_learn_1904_compiled.cpp - Detection result
components/edge_impulse/edge-impulse-sdk/classifier/ei_fill_result_struct.h
Inference runtime
- TFLite Micro
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro - Micro context
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro/micro_context.h - Tensor API
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/core/c/common.h
Embedded acceleration
Other components
- Device user
Connections 17
- Device user → Keyword app: starts detection
- I2S microphone → Keyword app: provides audio
- Keyword app → Audio AGC: adjusts gain
- Keyword app → Classifier API: runs classifier
main/keyword.cpp - Audio AGC → Classifier API: supplies audio
main/keyword.cpp - Classifier API → Audio features: extracts features
- Audio features → DSP math: uses DSP math
- Classifier API → Compiled model: runs model
- Classifier API → Model metadata: uses configuration
components/edge_impulse/edge-impulse-sdk/classifier/ei_fill_result_struct.h - Classifier API → TFLite Micro: invokes runtime
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro/kernels/conv.cc - TFLite Micro → Micro context: uses context
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro/micro_context.h - Micro context → Tensor API: uses tensor API
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro/micro_context.h - TFLite Micro → CMSIS-NN: uses kernels
components/edge_impulse/edge-impulse-sdk/tensorflow/lite/micro/kernels/conv.cc - TFLite Micro → ESP-NN: uses acceleration
- Classifier API → Detection result: formats results
components/edge_impulse/edge-impulse-sdk/classifier/ei_fill_result_struct.h - Keyword app → Status LED: signals detection
- Status LED → Device user: shows detection