MicroWakeWord custom V2 wake words: taking requests (12/2 only)

Configuration Voice Assistant
21

Some suggestions that I think would work well:

Hey Matey! (Or just ‘Hey Mate!’, if that could work)
Hey Homie! (Since this is Home Assistant, after all)
Marvin! (For the Hitchhiker’s Guide fans)
Bishop! (Aliens)
Cerebro! (X-Men)

22

Ive got it the .yaml is wrong configured
The original was:

{
    "type": "micro",
    "wake_word": "Okay Computer",
    "author": "Michael Hansen",
    "website": "https://www.home-assistant.io",
    "model": "okay_computer.tflite",
    "trained_languages": [
        "en"
    ],
    "version": 2,
    "micro": {
        "probability_cutoff": 0.97,
        "feature_step_size": 10,
        "sliding_window_size": 5,
        "tensor_arena_size": 30000,
        "minimum_esphome_version": "2024.7.0"
    }
}

and i changed it to:

{
  "type": "micro",
  "wake_word": "computer",
  "author": "Leland Olney",
  "website": "https://github.com/JohnnyPrimus/Custom_V2_MicroWakeWords",
  "model": "computer.tflite",
  "trained_languages": [
    "en"
  ],
  "version": 2,
  "micro": {
    "probability_cutoff": 0.66,
    "feature_step_size": 10,
    "sliding_window_size": 10,
    "tensor_arena_size": 22860,
    "minimum_esphome_version": "2024.7.0"
  }
}

also you need to use the RAW link to configure it.

micro_wake_word:
  models:
    - model: "https://raw.githubusercontent.com/oOJoshOo/Custom_V2_MicroWakeWords/refs/heads/main/models/computer/computer.json"

EDIT: It installs then, but the model doesnt load and crashes the ESP after that

2 Likes
23

This is excellent, would you consider “hello computer”. A great not to Star Trek 4.

24

The computer one doesn’t work.

:sleepy:

25

Sorry to “resurrect” this thread, but maybe you have an opinion/answer: if i only want a wake word to work for me anyway, and if i had the time to do so: wouldn’t it then produce better results if i recorded myself 1000 times over the stretch of a few days in my own apartment? Different times of day, different positions, different devices, different moods, etcetc

26

Your samples will be augmented into more by cahnging pitch and rirs and should work Ok its when you have 1000 initial samples that are not your voice that are this American English with very little variation it becomes overfitted to that American English voice if you want to overfit to your voice, yeah then do so and similary it will not work very well for anyone else unless equally similar, but for you if there augmentation and training is any good it should work quite well for you.
I did some models a long time back using my voice and fewer samples and they worked well.
ProjectEars/dataset/reader at main · StuartIanNaylor/ProjectEars · GitHub was a cli reader that collected wake word but I was using the same mic for sample collection and KWS.

A Wakeword model is just a dumb image classification model statistically comparing spectragraph/mfcc images to the ones being created by your voice throught the preprocessing algs in place on the device you use.
So the best samples are from the device of use as every mic device and alg has a signature that will just increase entropy and make things less accurate.
Also distance its pointless recording farfield samples as you can only detect RiRs (Room impulse response aka reverberation) by having multiple mics that will be able to detect the TDOA (Time difference of arrival). In a mono sample as sound bounces off all surfaces the further distances mix into the sample at longer times and reverberation at a mono source just becomes difference in the spectrograph again increasing entropy and lowering accuracy.
So just record broadcast like mic differences as RiRs are usually augmented afterwards.

27

Coco please! or Hey Coco