Home Assistant Voice PE device together with openWakeWord?!

sobchek · June 10, 2025, 3:25pm

Yes, ESPHome 2025.3 (i think) broke this.

Still trying to find a way around it that is compatible with the updates moving forward.

bvelte95 · June 26, 2025, 11:48am

Is there any update so far? Thanks a lot for your efforts!

sobchek · June 26, 2025, 2:04pm

I downgrade ESPHome to 2025.2. This seems to be the only way until the ESPHome team adds whatever resources are necessary to use custom wake words with MicroWakeWord. Its running smoothly with 2025.2 though.

bvelte95 · June 26, 2025, 5:08pm

I have to downgrade the ESPHome Addon, right? How about the builder app?

sobchek · June 26, 2025, 6:45pm

There is no official way to downgrade, unless you created a backup of the add-on at the time when you updated. I hadn’t, so I installed this custom repo into add-ons:

Its an archive of previous ESPHome versions, so I went with 2025.2, as that was the last version I recall it working correctly.

Install this then remove the official release ESPHome add-on. It will automatically find your devices. And no, the add-on is the builder app, same thing.

sobchek · July 3, 2025, 2:12pm

You can train a custom model for micro wakeword using this:

[GitHub - MasterPhooey/MicroWakeWord-Trainer-Docker](https://MicroWakeWord Trainer)

You will have to play with settings. But with this you can run the latest ESPHome Builder and HAVPE release. No hacks required.

NathanCu · July 3, 2025, 2:19pm

Dangit I had held off buying that EGPU until you posted that! RIP wallet.

sobchek · July 3, 2025, 2:21pm

Yes. Ran this yesterday and it works! Going to do a second run tonight with higher sample count. Will post any settings that make a noticeable difference.

NathanCu · July 3, 2025, 2:23pm

I was purposely waiting to do Friday’s wake word until I got through the core build. Yeahhh. Well… It’s time. I was hoping the builder was going to catch up. Looks like a good alternative if it’s CUDA only (the right choice). If someone has an IPEX alternative (falls down laughing) I’m all ears…

(starts spend justification doc for adding TB4-Nvidia EGPU. Ooooh a Blackwell rtx if we’re going there. Burn it down…)

bvavra1 · July 10, 2025, 8:38pm

How does this help bypass needing a custom build to run custom wake words on HAVPE? I had used the original hosted version of the microwakeword trainer to generate my wake word, which I uploaded to my Voice Assistant in HA, but I still cannot select it in the device with today’s version of the default HAVPE firmware.

(EDIT: To clarify, I had the original workaround up and running until I recently restored factory settings on my HAVAPE and ran into this new ESPHome breaking issue.)

NathanCu · July 10, 2025, 9:03pm

With this you basically take control of a HAVPE load only a diff of the production yaml that says yeah all that except this new microwakeword. (it doesn’t work with openwakeword you will need to compile a new one.) thays how.

Compile wake word. Put it in a git repo, take control of a havpe, drop in the right stuff to make it load the custom wake word and bobs yojr uncle.

And because it’s the diff method. If they ever push a new build you just recompile yours and it pulls new with your custom wake word.

The hardest part is getting a good wake word.

sobchek · July 10, 2025, 9:21pm

You are pulling the official config and overwriting or adding the necessary lines, as well as pointing to your local directory for the custom files

CyberMage · October 7, 2025, 12:48am

Two and a half months go by and that repo for training microwakeword, and the user, are gone from github.

sobchek · October 7, 2025, 1:30am

Yeah seems like the masterphooey docker has been rugged.

You can use this though, looks the same:

github.com/kahrendt/microWakeWord

notebooks/basic_training_notebook.ipynb

ef8befc49

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "r11cNiLqvWC6"
   },
   "source": [
    "# Training a microWakeWord Model\n",
    "\n",
    "This notebook steps you through training a basic microWakeWord model. It is intended as a **starting point** for advanced users. You should use Python 3.10.\n",
    "\n",
    "**The model generated will most likely not be usable for everyday use; it may be difficult to trigger or falsely activates too frequently. You will most likely have to experiment with many different settings to obtain a decent model!**\n",
    "\n",
    "In the comment at the start of certain blocks, I note some specific settings to consider modifying.\n",
    "\n",
    "This runs on Google Colab, but is extremely slow compared to training on a local GPU. If you must use Colab, be sure to Change the runtime type to a GPU. Even then, it still slow!\n",
    "\n",
    "At the end of this notebook, you will be able to download a tflite file. To use this in ESPHome, you need to write a model manifest JSON file. See the [ESPHome documentation](https://esphome.io/components/micro_wake_word) for the details and the [model repo](https://github.com/esphome/micro-wake-word-models/tree/main/models/v2) for examples."
   ]

This file has been truncated. show original

sobchek · October 7, 2025, 2:40am