Hacker News Reader

Comment by LawnboyMax | Hacker News Reader

I highly recommend Rasa Core (https://github.com/RasaHQ/rasa_core) if you are looking for an open-source virtual assistant. Very active and helpful community, and lots of input channels for integrating your assistant with messaging platforms (https://rasa.com/docs/core/connectors/).

I am not affiliated with Rasa, just had a really good experience developing a few projects with it.

View on HN

I'll +1 SNIPs or Rasa, they're both really nice. It looks like the NLU part of Leon is a logistic regression classifier (https://github.com/leon-ai/leon/blob/360d1020c4bd8bf1df37646...) so it's just doing intent detection, not any slot filling. Maybe someone can add calls to Rasa's HTTP API (https://rasa.com/docs/core/server/#) to integrate with Leon?

How well do these compare to Mycroft[1]?

[1]: https://mycroft.ai/

Haven't looked at MyCroft before. It looks like MyCroft exposes less of the nuts-and-bolts of modeling? I'm not sure where I would plug in a custom entity extraction or intent detection model, but I do see that it lets you add custom 'skills'.

Yes, this can all be handled with skills and the dictionaries that go along with skills. I don't even program much in python, but I've found it pretty simple to make add-ons for mycroft that have multiple entities for different intents.

I have had a Mycroft Mark I for about 6 months, and have a Mark II on order since I ordered them as a special bundle about a year ago. My feeling is Mycoft is aiming to be more of a stand alone product, closer to the base level echo or home as opposed to a framework to build things.

I've looked at some of the other projects, but one thing that appealed to me about Mycroft was they offered a complete hardware device. The Mark I is just a raspberry pi in a 3d printed case with custom lights, where as the Mark II is going to be more custom hardware (I think) whenever its finished. If you buy one of their devices you you could get it running just using their web interface and buttons without any knowledge of git, or ssh.

They have some ambitious plans, but their execution leaves something to be desired. The base level functionality is pretty good and has been reliable for me on a daily basis. It has a fall back to Wolfram Alpha and Wikipedia so you can ask it general questions like unit conversions, ages of famous people, weather and such. The results are no where near as polished as Alexa, but surprisingly responsive on random subjects.

My main problem has been with the plugins. I got it working with Kodi and Home Assistant and life was good, but as they have upgraded the project at various times these plugins have randomly stopped working for months at a time, and delving into why really drags you into more complicated troubleshooting. Other times it will randomly update and hang until you login via ssh and apt update the system to get it going again.

That said since they got their plugin marketplace working recently it has been pretty reliable and easier to get new plugins working.

You can build their product on your own, I believe you could even build the complete Mark I, but they have some odd "premium" options like donating 2$/month to get their newer voices.

They have a goal of privacy, but I believe the default processing for voice is Google, so big brother isn't far away. They are working, on a private backend https://github.com/MycroftAI/personal-backend and you could choose other voice backends so it should be pretty decent from a privacy perspective. At least good enough for me.

I've played with it a bit over the last year, and I can't help but assume I'm doing something wrong, but I just can't get it to work reliably. I installed it on a Raspberry Pi 3B+, took a 7-mic ReSpeaker mic array, and it only recognizes the wake word around 10% of the time, and then only if I speak really slowly and enunciate as if I were talking to a small child. I tried just recording audio from the mic and playing back to see if that was the problem, but it sounds crystal clear.

In contrast, a Google Home (which I really want to replace with something I trust more) sitting in the same location recognizes its wake word nearly 100% of the time.

Rasa seems to be text only though?

Or am I missing something?

Correct, Rasa is text-only. Core purpose of Rasa NLU (https://github.com/RasaHQ/rasa_nlu) is user text to intent translation (with entity extraction); for Rasa Core it is mapping user intent to assistant response . Speech-to-text is one layer above these and Rasa doesn't handle it.

The easiest way to use it with voice is to connect it to a platform that supports voice input.

Reply on news.ycombinator.com