A keyphrase detector (keyword spotter) that runs entirely on-device using a Google Coral Edge TPU. This is my fork of google-coral/project-keyword-spotter, the Google Research demo, kept in my account for Coral hardware experiments.
The concepts
Keyword spotting is the speech-processing task behind hotwords like “OK Google” — detecting a predefined short phrase in a stream of audio. The model in this repo recognizes about 140 short keyphrases (“move left”, “position four”, and so on) in a two-second audio window, with each output neuron of the network corresponding to one phrase. Because inference runs on the Edge TPU, detection is fast, local, and needs no cloud round-trip — the same privacy-friendly, on-device pattern I like for homelab voice control.
How to use it
Beyond the wrapper code, the repo ships three example programs:
run_model.py— prints the top detections with confidence scores each time inference runs.run_hearing_snake.py— a voice-controlled version of the game Snake.run_yt_voice_control.py— controls a YouTube player in a browser tab with phrases like “next video” or “mute”.
Running it
You need a Coral Dev Board or USB Accelerator set up with the current runtime. Then:
sh install_requirements.sh
python3 run_model.py
Speak a supported keyphrase and you’ll see it recognized in the console output with its confidence.
Source code: github.com/avnit/project-keyword-spotter Upstream: github.com/google-coral/project-keyword-spotter