TinyML keyword spotting on XIAO ESP32S3 Sense

What you’re building

Seeed’s tutorial builds an offline keyword-spotting model for the XIAO ESP32S3 Sense using its expansion board’s PDM microphone and Edge Impulse. Audio is recorded as 16 kHz, 16-bit WAV files to a FAT32 microSD card, divided into one-second labeled samples, transformed into MFCC features, and used to train a small convolutional classifier. Edge Impulse exports a quantized Int8 Arduino library that runs inference on the ESP32-S3; the sample outcome lights the onboard LED when the trained “hello” class is detected. Required hardware includes the Sense expansion board, a data-capable USB-C cable, and a microSD card no larger than 32 GB. The documented optimized deployment is explicitly non-EON and requires the cited preliminary ESP-NN replacement; label order and generated library name must also be updated in the example sketch.

Key steps

1. Attach the Sense expansion board, format a microSD card of 32 GB or less as FAT32, and enable OPI PSRAM.
2. Upload the matching ESP32-core WAV recorder and capture spaced examples for each keyword plus noise or unknown classes.
3. Upload recordings to Edge Impulse, split the ten-second files into one-second labeled samples, and retain a train/test split.
4. Create a one-second impulse with MFCC processing and a Keras classifier, then generate features and train the model.
5. Build a quantized Int8 Arduino Library with EON disabled and apply the documented ESP-NN replacement.
6. Install the generated ZIP library, update its include and label mapping in the inference sketch, upload, and test live keywords.