Conclusion

A conclusion of the project and the Tech-Stack

Technology Documentation Installation Usability Accuracy/Usefulness Recommendation
Precise good easy good ok yes (restricted)
Kaldi good tricky good good yes (restricted)
FuzzyWuzzy - very easy good bad no
MaryTTS good ok ok understandable yes
HermesMQTT good - good full control yes

MyCroft-Precise

Mycroft-Precise is well documented and easy to install.
You should not install it on your Raspberry Pi, but on a more powerful computer due to its performance and storage size.
The training is easy and the model is easy to use, but you have to invest some time to find a good wake-word.
You should install a tool to filter background sounds to increase the accuracy.
It is not possible to activate our voice assistant, while music is playing.

Yes, we would recommend it under these restrictions!

Kaldi

The documentation of Kaldi is not very appealing, but extensive.
It is a bit tricky to install, but doable if you attentively and patiently follow the instructions.
After you installed it, it is very precise and fast.
It comes with a pretty well pre-trained base-directory, which is easy to extend.
In our case it was not possible to detect words, that are not stored in the sentences.ini of Rhasspy.
So it is not possible to add a “searching”-functionality to use sites like “Wikipedia”.

Yes, we would recommend it under these restrictions!

FuzzyWuzzy

FuzzyWuzzy is a pretty simple tool.
You do not really have to install anything, you can simply select it in Rhasspy.
It is very limited in the amount of intents and sentences it can process and the spoken Sentences should match these.
Even if there is not a big amount of sentences it had some issues to recognise the right intents.

No, we would not recommend it!

MaryTTS

The installation of MaryTTS was a bit confusing due to the multiple tools you can install (e.g. the “GUI”-version or the “Installer”-version).
To understand, which version is the best for your purposes, you should read the well written documentation.
The voice we have chosen was understandable, but had some words with misspelling and some problems with negative numbers.
These problems are preventable and not really a counterargument.

Yes, we would recommend it!

HermesMQTT

HermesMQTT is easy to use if you know how to work with MQTT and (in our case) Node-Red.
You do not have to install something, just select it in Rhasspy and use the HermesMQTT-topics to set up your dialogue-manager.
We used the documentation of Snips to learn how to use the topics and it was easy to implement.
You should use HermesMQTT, if you want to have full control over you voice assistant.

Yes, we would recommend it!

What’s next?

Here you can read how we imagined the user-experience at the beginning.