Trigger word – GRU

I am just finishing the last exercise from NG – this time it’s ‘Trigger word’. It works by saying a key word and the smart network recognises it and perhaps activates some device. The network starts with a 1D convolution, a spectrogram, 2 layers of GRU, and finally, a dense network is added, capped off with a classic sigmoid. In total, it’s only about half a million parameters. I tried uploading my own voice commands into the exercise, and surprisingly, it understands my English too :) From today onwards, the networks will literally listen to me word for word!