Custom wake words for ESP32 and microcontrollers
microWakeWord is a free, open platform that lets you train quantized INT8 TFLite wake word models optimized for ESP32-S3, ESPHome, and Home Assistant Voice PE. Detection runs entirely on-device with no cloud, no subscription, and no privacy compromise.
How it works
Pick a wake word, generate a training dataset from multilingual text-to-speech voices, and launch a GPU training job on our serverless infrastructure. A few minutes later you receive a ready-to-use .tflite model plus the JSON manifest required by ESPHome's micro_wake_word component.
What you get
Every trained model ships as an ESPHome-compatible package: the TFLite file, a manifest with probability_cutoff and sliding_window_average_size, and a README explaining exactly how to wire it into your ESP32-S3 device. Models are also published to a public library so the community can browse, download, and rate them.
Target hardware
Models are sized and quantized for the ESP32-S3 family used by Home Assistant Voice PE, ESP32-S3 Box 3, Seeed reTerminal DM, and most ESPHome-based voice satellites. They run in roughly 35-45 kB of tensor arena at 16 kHz with a 20 ms feature step size.
More
This page is an app that requires JavaScript. Enable JavaScript to use the training center and interactive library.