Wake Word Engine


We decided to go with Mycroft-Precise for the given reasons and started to train our wake-word.
Hereinafter we explain, how we made it and what the difficulties were. Mycroft-Precise uses .wav files to train a model, which then can be used by Rhasspy to activate the voice assistant.
From the recorded sound waves we can then use “Text to Speech” (TTS) and “Intent Recognition” to control other devices or ask for news and weather.

How to find a wake-word

Our wake-word should be about six to eight phonemes.
If it is too short it is hard to detect or there might be more false positives, because our wake-word might be contained in commonly used vocabulary.
If it is too long it is harder to activate the assistant and less enjoyable to use.
We also want different phonemes since they create an easy distinguishable sound signature.

Since we do not have a brand and do not want to promote one, we turned to mythologie and found “Heimdall”.
Heimdall is associated as possessing foreknowledge, keen eyesight and hearing, and keeps watch for invaders (wikipedia).

How to train your own wake-word?

This is a shorted version of the original Documentation.
After you installed Precise by following this step-by-step instruction,
you are able to use source .venv/bin/activate to activate the virtual console and the following commands:

  • precise-collect to record own .wav files
  • precise-train to train a model with given sound-files initially
  • precise-train-incremental to reduce false activations at trained model
  • precise-test to test your model with prerecorded soundfiles
  • precise-listen to test your model live with a connected microphone (in our case the ReSpeaker 4-Mic Array)
  • precise-convert to convert your .net-Keras model to a .pd-TensorFlow model

Recording and training

We recorded some audio samples using the precise-collect-command and moved most of these .wav-files into the mycroft-precise/heimdall/wake-word/-directory
and a few into the mycroft-precise/heimdall/test/wake-word/-directory.
We collected some files recorded by our friends and family, too.
After we sorted all recorded samples into the right directories, we started the training using the precise-train-command.

    precise-train -e 60 heimdall/

Testing and improving

After the first training we tested our model using the precise-listen-command and noticed, that there are many false detections.


To reduce these false detections we downloaded the Public Domain Sounds Backup as recommended in the original Precise documentation.
To do this, we downloaded the package to the data/random-directory and unzipped it.

    cd data/random
    7z x pdsounds_march2009.7z
    cd ../..

To convert all these downloaded .mp3-files to the needed .wav-format we used the following script:

    for i in $SOURCE_DIR/*.mp3; do echo "Converting $i..."; fn=${i##*/}; ffmpeg -i "$i" -acodec pcm_s16le -ar 16000 -ac 1 -f wav "$DEST_DIR/${fn%.*}.wav"; done

Now we had all files in the right format and were able to train the model again. To do so, we used the precise-train-incremental-command, which takes clips from the data/random-directory, copies these to the heimdall/not-wake-word-directory and retrains the model. This process lasts a few hours on the raspberry pi 3B+.

    precise-train-incremental heimdall/ -r data/random/

Directories can be different to your setup.

You can repeat this whole process until you are happy with the result of the wake-word-detection.

Use your model in Rhasspy

Before you can use your model with Rhasspy you need to convert it from the Keras .net-format to the TensorFlow .pb-format. We have used the precise-convert-command to do so.


The command creates the following three files:

  • heimdall.pb
  • heimdall.pb.params
  • heimdall.pbtxt

After we copied all these files to the rhasspy/profiles/de/precise-directory, we were able to select and use them in the Rhasspy-GUI on http://<hostname>:12101.
Alternatively you can add the following lines to your profile.json.

          "precise": {
              "model": "heimdall.pb"
          "system": "precise"



What’s next?

Precise triggers Kaldi, to listen to our commands.