Home Assistant has its own voice assistant called Assist. It comes with some pre-trained wake-up words that you can use for voice commands, including “Okay Nabu”, “Hey Jarvis” and “Hey Mycroft”. I wanted to create my own custom wake word and it was easier to do than I expected.
Teaching the waking word is easier than it looks
No prior knowledge or powerful equipment is required
I always assumed that learning my own custom wake word would be a process that required some pretty powerful hardware and would take hours to complete. In fact, you can do all the processing through the browser without any of the processing happening on your device. Although it’s a bit tricky, you can theoretically do it on your phone.
The whole process goes to the island Google Colab notebook Powered by the Home Assistant team. This is a web page that contains all the code and instructions you need to create a custom wake word with its code It runs on Google’s computers to his own place.
The first section is where you enter your custom wake word. Running this section installs the required tools and then creates your oral version waking word. If it doesn’t sound right, you can repeat the process using phonetic spelling until the generated wake word sounds the way you want.
The next section loads audio samples of things like background noise, music, and room reverberation. These are used to make the examples in which the particular wake word will be taught as realistic as possible. It takes about 15 minutes to download everything.
The last part trains the model. It creates multiple audio files of your custom wake word spoken with various background sounds added. These audio files are then used to train your model to recognize the word that is evoked even with background noise.
In general, the process took about an hour using the default settings. Once done, you end up with your custom wake word saved in two different formats: TensorFlow Lite and Open Neural Network Exchange (ONNX).
- Dimensions
-
84x84x21 mm
- Weight
-
96 g
The Home Assistant Voice Preview Edition is the first privacy-focused smart speaker built as an alternative to the likes of Amazon Alexa and Google Nest Mini. It adds voice assistant capabilities, including local-only processing, to a smart home powered by Home Assistant.
Adding a wake word to your home assistant is simple
Drop the model into the folder and restart openWakeWord
I wanted to create a catchphrase to make me feel like David Hasselhoff from Knight Rider. He could talk to or through his AI car KITT touch screen in the car or through his watch I used Google Colab to teach him the word “Hey KITT”.
I had to install it first openWakeWord plugin To the Housekeeper. This is an app that makes Home Assistant listen to wake words. I threw it away after it was installed hey_kitt.tflite file into Home Assistant. I used it Samba plugin to add it /share/openwakeword directory.
Then I set up a new voice assistant in Home Assistant. In the three-dot menu Add helper there is a window Add a wake word choice. Choosing this allowed me to choose openWakeWord as a wake-up word engine and Hello KITT it was possible to choose from a list of supported wake words.
My smart speaker couldn’t easily use the specific model
I tested wake words on my Mac instead
i have Smart speaker for Assist built using ReSpeaker Litebut it uses wake word detection on the device, so my openWakeWord model was not compatible. To test my wake word, I decided to install an old iMac voice satellite.
My ancient iMac wasn’t compatible with the newer ESPHome-based audio satellite options, so I had to install it using the outdated Wyoming Protocol integration. After a bit of fiddling I was able to launch Help by saying “Hey KITT” and stuff like that. turn the lights on and offhowever, I couldn’t get audio responses to work, so I couldn’t hear confirmation sounds or spoken responses from Assist.
I then tried again on an M2 MacBook Air, which can run Linux Voice Assistant to allow me to use my Mac as a voice satellite. Using the LVA, I was able to hear Assist’s sounds and responses through the MacBook’s speakers.
The results were pretty solid
I finally felt like David Hasselhoff
The Google Colab setup has default settings for training, but says that using more samples and training steps will improve the model. However, it will take more time to do so. Using the default settings, I was pretty sure the wake word wouldn’t be very effective. I was pleasantly surprised to find that it worked better than I expected.
Saying “Hey KITT” worked almost every time, and even with music playing, the wake word was recognized. I could say “Hey KITT” and ask for a light to turn off or some other smart home command and it would happen within a second or two. I finally knew what it was to be Michael Knight; now I just need to vote for Assist William Daniels.
It’s always been weird to say “Okay Nabu”.
Alexa isn’t the greatest wake-up word, but “Okay Nabu” has always felt like an odd choice. It’s nice to be able to create your own wake word, although “Hey Jarvis” is better. It’s much easier to do than I expected.
