Show HN: Dexter, a Voice Controlled Assistant
github.comI have to admit, voice assistants don't interest me much, but this looks like a pretty cool start to something that seems to be needed and lacking in the realm of voice assistants.
I'd like to see this gain traction and grow. Just because of the amount of people that seem to use and like voice assistants it's great to see some work on the open source, non-proprietary cloud side of things.
I also have to admit, I find most of thr current voice assistant technology vaguely creepy knowing every command is possibly stored and saved somewhere outside my control. The fact the systems that do this are constantly listening for phrases and of course the inevitable use of your data for marketing that these services are based on.
I understand the choices seem to be either pay for it or pay with your privacy and data, but I still hold out the naive hope for a technology based future focused on empowering individuals and not service providers and large corporations, sadly it seems this mostly comes from the open source world, which as many articles on hn have touched on, don't tend to get the support they need or they get picked up by large entities who take what they need and surround them by proprietary blobs.
Sorry this got a bit rambling, but anyway, this is a cool project, I hope to see it grow and gain support and adoption.
>I understand the choices seem to be either pay for it or pay with your privacy and data
Where have you seen those choices? What I've seen is either pay with your privacy and data (Cortana for example) or pay for it and at the same time surrender your privacy and data (Alexa, Google Nest).
Siri is more of a pay for it and keep the privacy/data.
No it isn't. See: Apple in China
If you were going for quality discussion you’d say “it isn’t in China”
I think you'd rather say that Apple will dress themselves up to look like what their audience wants to see - that they don't really care about privacy, but more the perception that you have it. Their actions in China back this view up. You see, the expectation in the West is that you make sure you look like you're doing the "right thing", but behind the scenes you accommodate state actors doing what they want with your data.
Or rather that they have a hierarchy of morality in which privacy is significantly higher than any other major tech company, and yet are willing to sacrifice it to fit local laws if the market is large enough.
Unless you have proof they are actively exposing your Siri data behind the scenes in the west, your point is simply fear mongering.
They have invested and spent beyond-significant resources in a wide array of world best privacy technology - all the way down to the chip level; and all the way up to their entire structure of the way they do machine learning, including many novel papers and a lot of work into maturing the field of distributed, differential, private ML. It’s well beyond simple marketing. They also are actively making huge and risky moves that upset large competitors and partners, ongoing, re:privacy. If they were just paying lip service to this, they’d need not to do even a tenth of what they’ve done. To try and claim they aren’t is ridiculous, especially when literally no one else even remotely comes close.
In fact, there exist no platforms you can use today that would come near the level of highly scrutinized, tested and invested in top to bottom security in privacy tech than Apple devices. It’s not even a contest, and that includes all open firmware/stacks.
Sorry I guess I should have meant theoretical choices...at this point there isn't much of a choice.
In response to a few things in your comment:
I was actually in the same opinion as you about remote storage and always on wake detection, and I actually found out that it's possible to make one without all of this!
There are "offline wake work detector libraries" like Snowboy.ai that prevent one of those things, and you can also get self hosted versions of even popular language parsers like the Azure Speech Recognition that you can run in a docker image and have everything stored locally (or not at all!)
Not suggesting you have to go make your own. It just makes me hopeful for the future where we can build useful assistants that don't need all this privacy l-destroying crap in them :)
I wish somebody made a text-based assistant that has similar functionality. As a non-native English speaker I don't think voice assistants will ever become accurate enough to be usable by me.
I use voice2json on my homebrew RPi voice rig, and it has a plain-text input mode [0]. With a little Node-RED + MQTT + shell magic, it's pretty neat/convenient to type into the terminal: `house "turn on lamp"`. :)
Dexter has a couple of different input systems: PocketSphinx and DeepSpeech.
PocketSphinx seems to have better language coverage but lower accuracy than DeepSpeech, which only has English and Mandarin Chinese.
DeepSpeech was not great about 2 years ago when I first started Dexter and, being a Brit, I had to fake an (awful) American accent to make it work. It's come on well in that time though; I can use my own voice now, which is a relief to those near by. No clue how good its Mandarin recognition is however.
Other folks in surrounding threads have flagged Kaldi which I'll try to take a look at as well (though it also looks to be only English and Mandarin too).
I didn't try it out yet but Kalliope is a similar project (RPi voice assistant) but has an API where you could input a text order via HTTP:
https://kalliope-project.github.io/kalliope/api/synapses/#ru...
How is this different from previous voice assistant projects designed for the raspberry pi. Projects such as: Rhaspy, Mycroft, and Jasper?
It was mainly a toy project for me; a bone to chew on if you will. I wanted something which I could mess with easily myself and which was heavily pluggable.
Integrating it with various external peripherals should be easy and it's designed to run completely on its own (i.e. no cloud for a lot of things).
Nice one but how is it different than so many other Raspberry Pi based voice assistants?
I added a "Related work" section to the README. It's a tad vacuous since, really, these things are all pretty similar. Dexter can do things with swirly lights though. Swirly lights are cool.
Cool project! Is the name perhaps inspired by either: 1) The kids cartoon scientist with a lab, or 2) The Miami homicide bloodsplatter specialist moonlighting renegade serial killer?
It's actually a pun: Dexter, as in the Latin for "right", as in your right hand. Bad I know.
If it's the 2), the assistant should be called the Dark Passenger.
Having the voice of the cartoon would attract me though, for the fun factor. Imagine having Dee Dee for error messages:
User: "Dexter, play some music."
Dee Dee: "Ah, I don't want to."
Dexter: "Dee Dee! What did you do to the network connection! I can't reach the Internet! We're doomed!"
This looks great so far. I'm gonna play with this soon.
I ask this about every voice assistant project: Is the wakeword changeable, or can I set it to multiple?
Yes and yes.
This is awesome! I've been looking for something like this for a while- I'll have to try it when my Pi arrives.