The biggest problem to solve is parsing generative responses in the domain of mathematics with LaTeX, or something similar. Try making a conversational math teacher, and experience this directly. Most APIs don’t do this, and the AI just pronounces math terms phonetically, and not in standard form or with correct pronunciation. Writing a LaTeX parser is a a non-trivial task, and the API returns are not contextually aware of when they’re speaking math, so those flags don’t even exist in the XML or JSON replies to the API calls. This is a major barrier to creating effective and reliable AI agents for teaching math and science.
We have the ability now to add phonetics for hard to pronounce words or in-game words. For example, Cygnus Enterprises had a resource named Cygnite Ore and it was made up. The AI couldn't recognize the word or pronounce it at first. They were able to connect the phonemes with it and set a preference so that the AI was biased to recognize that word over other possible words. Your use case would be possible but you would have to set up all the words you wanted to be pronounced a certain way with the phonemes. We currently are prioritizing video game use cases but that could change in the future!
@@inworldai Getting closer! I have been working with NVIDIA to see if I can leverage something like Riva or ACE with my Inworld Metahumans. I have to solve this problem and make my proof of concept work before we commit to an Enterprise license. It’s a tough nut to crack, unfortunately. Not too long ago, I was contacted by an independent company who are making AI trainers, and they are facing the same problem. And they’re an entire team! I’m just one person, a full-time professor and advisor with no support staff, trying to get this done on my own. It ain’t easy!
okay this is actual pretty darn interesting, love to see more in the future!
i searched this video not only cause i love the wol experience, but the intro music is just soothing
The biggest problem to solve is parsing generative responses in the domain of mathematics with LaTeX, or something similar. Try making a conversational math teacher, and experience this directly. Most APIs don’t do this, and the AI just pronounces math terms phonetically, and not in standard form or with correct pronunciation. Writing a LaTeX parser is a a non-trivial task, and the API returns are not contextually aware of when they’re speaking math, so those flags don’t even exist in the XML or JSON replies to the API calls. This is a major barrier to creating effective and reliable AI agents for teaching math and science.
For example, if we have a component such as Ax^2, it’ll usually be pronounced as “axe two”. This is the opposite of helpful.
We have the ability now to add phonetics for hard to pronounce words or in-game words. For example, Cygnus Enterprises had a resource named Cygnite Ore and it was made up. The AI couldn't recognize the word or pronounce it at first. They were able to connect the phonemes with it and set a preference so that the AI was biased to recognize that word over other possible words. Your use case would be possible but you would have to set up all the words you wanted to be pronounced a certain way with the phonemes. We currently are prioritizing video game use cases but that could change in the future!
@@inworldai
Getting closer! I have been working with NVIDIA to see if I can leverage something like Riva or ACE with my Inworld Metahumans. I have to solve this problem and make my proof of concept work before we commit to an Enterprise license. It’s a tough nut to crack, unfortunately. Not too long ago, I was contacted by an independent company who are making AI trainers, and they are facing the same problem. And they’re an entire team! I’m just one person, a full-time professor and advisor with no support staff, trying to get this done on my own. It ain’t easy!