Machines Parlantes
September 27, 2022 8:17 PM Subscribe
Quazel lets you practice languages by conversing with a machine learning system.
So far this seems like it could be cool though I keep talking to it and then discovering that it isn't recording. And when it is recording, the actual recognition of what I am saying is very poor, which may be a result of my shitty accent, but may just be bad language recognition.
Still, it is very unintimidating to talk to this thing, and I need a lot of practice with that, so anything that gets me speaking more is good.
posted by jacquilynne at 5:53 AM on September 28, 2022 [3 favorites]
Still, it is very unintimidating to talk to this thing, and I need a lot of practice with that, so anything that gets me speaking more is good.
posted by jacquilynne at 5:53 AM on September 28, 2022 [3 favorites]
Interestingly, when you use its grammar checker on your sentences, you can make all the changes it suggests, and run it again, and it will iteratively improve things. I think the first time, it fixes the basic errors in what you wrote and then if it doesn't find anything, it will look for word choice improvements. Sometimes it changes things it suggested the first time around, even.
posted by jacquilynne at 6:23 AM on September 28, 2022
posted by jacquilynne at 6:23 AM on September 28, 2022
Self-plug: Osio (https://www.edom-electric.com/osio) has web and phone versions, corrections and suggestions, and allows you to select from a broad set of useful topics.
posted by thandal at 6:54 AM on September 28, 2022
posted by thandal at 6:54 AM on September 28, 2022
Now I'm trying the web app for Osio, but it only ever recognizes the first word of a sentence. I'm finding the interface much less clear generally. Possibly it is more powerful -- it seems to have more going on in the interface -- but it isn't super apparent to me what I can do at any given moment, or how corrections work. Maybe I will try it on mobile and see if it is more intuitive there.
posted by jacquilynne at 7:18 AM on September 28, 2022
posted by jacquilynne at 7:18 AM on September 28, 2022
This is kind of amazing. The bot opened one of the lessons with "What brings you to the park today?" and when I told them that we're not in the park we're in jail they explained that they were here to visit their husband. Follow-up questions revealed that he was doing two years for stealing a car, but the bot was being pretty casual about it. I said, "Two years for stealing a car? That's too much!" and the bot responded that it's been pretty hard on the family but if you commit a crime you have to pay for it.
posted by charismatic megafauna at 8:20 AM on September 28, 2022 [8 favorites]
posted by charismatic megafauna at 8:20 AM on September 28, 2022 [8 favorites]
^ so what I hear you saying is I can cancel all my streaming services and just engage with this thing
posted by elkevelvet at 10:46 AM on September 28, 2022 [3 favorites]
posted by elkevelvet at 10:46 AM on September 28, 2022 [3 favorites]
It did well with the language I'm currently studying (Spanish), then wanted me to log in.
posted by doctornemo at 11:10 AM on September 28, 2022 [1 favorite]
posted by doctornemo at 11:10 AM on September 28, 2022 [1 favorite]
I'm impressed. I got pretty far into convincing a train agent to take a bribe in diamonds in exchange for a ticket to the moon before the account barrier came up. Perhaps it's cheating to pick a scenario where a human would normally pretend not to know what's going on. . . but, it worked surprisingly well. (I'll probably make an account, 'cause this could actually be useful. . . but, I wish they wouldn't do that.)
posted by eotvos at 12:25 PM on September 28, 2022 [4 favorites]
posted by eotvos at 12:25 PM on September 28, 2022 [4 favorites]
This is very cool. Great use case for GPT-type technology.
posted by dusty potato at 4:55 PM on September 28, 2022
posted by dusty potato at 4:55 PM on September 28, 2022
Although, I will say, while it seems quite unlikely that user responses are directly training the model as wondered above, I do feel a little uncomfortable that I can't find any sort of privacy policy re how your voice clips are used. I can see user input here being a very valuable commodity for some other machine learning use.
posted by dusty potato at 4:58 PM on September 28, 2022
posted by dusty potato at 4:58 PM on September 28, 2022
very valuable commodity for some other machine learning use
Usually the most valuable application of gathered data is in refining the service you're using. eg, it's really common to train a system initially on some big public datasets (like CommonVoice), and see that it works OK, but not great... And then you refine naturally using user inputs. In this case, you've got lots of 'foreign' accents and poor pronunciation making the voice recognition harder, and a nice easy way for people to correct what the machine heard, which very easily becomes new training data, perfectly adapted to people's actual usage. The same data could be useful for other systems, but most other applications won't have /quite/ the same degree of early-learning speakers.
Turning back conversations for training data for the text-generating part of the model is more dangerous; if you train on the language learner's conversations, you'll get bots that sound like language learners... I would guess they're using things closer to off-the-shelf for that part of the system and for the text-to-speech system.
posted by kaibutsu at 8:37 PM on September 28, 2022
Usually the most valuable application of gathered data is in refining the service you're using. eg, it's really common to train a system initially on some big public datasets (like CommonVoice), and see that it works OK, but not great... And then you refine naturally using user inputs. In this case, you've got lots of 'foreign' accents and poor pronunciation making the voice recognition harder, and a nice easy way for people to correct what the machine heard, which very easily becomes new training data, perfectly adapted to people's actual usage. The same data could be useful for other systems, but most other applications won't have /quite/ the same degree of early-learning speakers.
Turning back conversations for training data for the text-generating part of the model is more dangerous; if you train on the language learner's conversations, you'll get bots that sound like language learners... I would guess they're using things closer to off-the-shelf for that part of the system and for the text-to-speech system.
posted by kaibutsu at 8:37 PM on September 28, 2022
when I tried it it just asked me what my typical day was like & I said I played a lot of video games & it asked me how many hours & I said well I had to play them for work so 10 jikan gurai & it was like 10 jikan?!?! that is SO MUCH, are you on break?
& I was like CHIGAIMASU YO! I play them for work I said & the correction thing was like "do you mean 'I don't play them for work I said'" & I was like "I'm pretty sure that's not what I mean" so I went ahead with "shigoto de purei suru to iimashita" which like maybe the "de" is incorrect there but at this point I felt like the thing did not understand me on a fundamental level & I hit enter & it was like "oh okay" & then immediately asked me to make an account & in my head in English I was like "YOU NEGGED ME ABOUT MY JOB AND NOW YOU WANT TO GET MARRIED"
posted by taquito sunrise at 10:13 AM on September 29, 2022 [2 favorites]
& I was like CHIGAIMASU YO! I play them for work I said & the correction thing was like "do you mean 'I don't play them for work I said'" & I was like "I'm pretty sure that's not what I mean" so I went ahead with "shigoto de purei suru to iimashita" which like maybe the "de" is incorrect there but at this point I felt like the thing did not understand me on a fundamental level & I hit enter & it was like "oh okay" & then immediately asked me to make an account & in my head in English I was like "YOU NEGGED ME ABOUT MY JOB AND NOW YOU WANT TO GET MARRIED"
posted by taquito sunrise at 10:13 AM on September 29, 2022 [2 favorites]
« Older The Gran Turismo of Animal Crossing | insects and rodents seem apparently never to enter... Newer »
This thread has been archived and is closed to new comments
posted by jackbishop at 2:55 AM on September 28, 2022