Create your own wake word

You can now create your own wake word to use with Home Assistant. The procedure below will guide you to train a model. The model is trained using voice clips generated by our local neural text-to-speech system Piper.

Want to know more about how this all works? Check out the openWakeWord project by David Scripka.)

Depending on the word, training a model on your own wake word may take a few iterations and a bit of tweaking. This guide will take you through the process step by step.


To create your own wake word

  1. Think of a wake word.

    • A word or short phrase (3-4 syllables) that is not commonly used so that it does not trigger Assist by mistake.
    • Currently, only wake words in English are supported.
  2. Open the wake word training environment.

  3. In section 1, enter your wake word in the target_word field. Enter wake word in target field

  4. In the code section next to the target_word, select the play button. The first time this can take up to 30 seconds.

    • If the play button does not appear, make sure your cursor is placed in the target_word field. Select play button

    • If it still does not show up, in the top right corner of the document, make sure it says Connected.

      • If it is not connected, select Connect to a hosted runtime. Connect to hosted runtime
    • Result: The pronunciation of your wake word is being created.

      • Once it is finished, at the bottom of the section, you see an audio file. Listen to it.

      Listen to demo of your wake word

  5. If the word does not sound correct to you:

    • Follow the instructions in the document to tweak the spelling of the word and press play again.
    • The word should sound the way you pronounce it.
  6. Once you are satisfied with the result, in the menu on top of the screen, select Runtime > Run all.

    • This will take around an hour. Feel free to do something else but make sure to leave the browser tab open. Runtime: run all
    • Result: Once this process is finished, you should have 2 files in your downloads folder:
      • .tflite and .onnx files (only .tflite is used)
  7. Congratulations! You just applied machine learning to create your own wake word model!

    • The next step is to add it to Home Assistant.

To add your personal wake word to Home Assistant

  1. Make sure you have the Samba add-on installed.
  2. On your computer, access your Home Assistant server via Samba.
    • Open the share folder and create a new folder openwakeword so that you have /share/openwakeword.
  3. Drop your shiny new wake word model file (.tflite) into that folder.
  4. Go to Settings > Voice assistants.
    • Either create a new assistant and select Add assistant.
    • Or, edit an existing assistant.
  5. Under Wake word, select openwakeword.
    • Then, select your own personal wake word.
    • If there is no Wake word option, make sure you have the add-on installed and successfully completed the $13 voice assistant for Home Assistant tutorial.
  6. Enable this new assistant on your ATOM Echo device.
    • Go to Settings > Devices & Services and select the ESPHome integration.

      • Under M5Stack ATOM Echo, select 1 device.
    • Under Configuration, make sure Use wake word is enabled.

    • Select the assistant with your wake word.

      Select the assistant with your wake word

  7. Test your new wake word.
    • Speak your wake word followed by a command, such as “Turn on the lights in the kitchen”.
    • When the ATOM Echo picks up the wake word, it starts blinking blue.


Troubleshooting wake word recognition

  1. If the ATOM Echo does not start blinking blue when you say the wake word, there are a few things you can try.
  2. Go to Settings > Devices & Services and select the ESPHome integration.
    • Under M5Stack ATOM Echo, select 1 device.
    • Under Controls, make sure Use wake word is enabled.
  3. If this was not the issue, you may need to tweak the wake word model.
    • Go back to the wake word training environment.
    • In section 3 of the document, follow the instructions on tweaking the settings and create a new model.

Troubleshooting performance issues of the training environment

The environment on the Colab space runs on resources offered by Google. They are intended for small-scale, non-commercial personal use. There is no guarantee that resources are available. If many people use this environment at the same time or if the request itself uses a lot of resources, the execution might be very slow or won’t run at all.

It may take 30-60 minutes for the run to complete. This is expected behavior.

Things you can try if the execution is very slow:

  1. Free of charge solution: This environment has worked for all the wake word models that were trained to create and test this procedure. There is a good chance that it will work for you. If it does not, try training your model another time. Maybe many people are using it right now.
  2. You can pay for more computing resources: In the top right corner, select the RAM | Disk icon.
    • Select the link to Upgrade to Colab Pro.
    • Select your price plan and follow the instructions on screen. Connect to hosted runtime