Wake word detection

Introduction

There are many examples of speech control devices in today's homes, Amazon's Alexa being a prominent example. These devices wait for a wake word, "Alexa" in case om Amazon's system. Only when this wake word is detected the following speech commands are sent to a server over the internet which will then interpret the command itself.

In the micro-speech example provided in the examples folder of tflite-micro the words yes and no are detected. To make the example work we need a microphone, converting the sound into an electrical signal and an ADC to convert this signal into numerical sound samples. The speech can then be recorded and save as .wav files, which can be used to train a neural network.

In the example the sound samples are fed into the neural network model which performs an action when either yes nor no is detected (e.g. printing out the result and lighting some lamps.

The microphones

Micropython already include an I2S driver, written by Mike Teachman, which can read digital I2S microphones. Mike also provide I2S example programs showing how to use his driver.

For the implementation of the examples we can use two different I2S microphones supported be the I2S driver:

The INMP441 MEMS I2S microphone

The hardware connections to the WeMos D1 bus are as follows:

-- Uli Raich - 2022-02-20

Comments