I have a startup idea that I'm not going to execute. I currently live with a senior social worker, and my mother was a social worker. Over the years have been deeply troubled by what I've heard. I think this tool could help them, even if it's a bit creepy.
Social workers in the UK have a big problem. Different teams, professions (police, medical, housing), deal with the same people, and often collect all the pieces required to understand and prevent a tragic situation. But they can't put them together.
There is an obsession with interpofessional meetings to tackle this, but every serious case review shows that the knowledge was there but not put together, leading to dead children. My perception is that often knowledge is mentioned, but when discussing a different related case, or isn't formally written down and passed to the right person.
My solution is inspired by the part of the CALO project that dealt with capturing and exploiting knowledge from meetings. There are some great ideas. If you do this, read up on it.
The product records all meetings, using existing meeting room Teleconference hardware. The meetings are put through automatic speech and speaker recognition on a best effort basis to produce a searchable index for that meeting. So you get something like a transcript and can skip to the point where Mr A is mentioned.
Supervisors, who in that field are criminally liable for mistakes by subordinates, already review minutes for meetings. Now they can search for any mention of Mr A, and see in what context he was mentioned and listen to that part of the meeting. Obviously there would be control of who could listen to what meetings, but these people already have big intrusive databases, so I don't think they are too worried.
Ideally, you could then build graphs from there of who is mentioned together, and do some Palantir style stuff to spot associations between people.
But the start is, very much, implementing a meeting transcriber by sowing together a few brands of speech recognition engine, a web interface for administrating and searching meetings, and selling it into social services.
If you are relying on teleconference hardware made by other companies, I suspect that you will have those companies cloning your software for their hardware.
I mean, I assume Cisco Telepresence stuff doesn't already record/transcribe meetings... If your idea starts to take off, I would bet that it will for sure in the near future. That seems like exactly the sort of functionality you would expect from that sort of hardware.
Well CALO yielded SIRI, as in on the iphone. That project had a number of parts, one was an AI personal assistant, another was meetings, and another was some kind of integrated battle control center.
Anyway, it yielded a lot of high grade speech recognition research and AI research, which companies are battling to license. If you recall Yahoo's weirdly overvalued purchase of Sumly, that was part of a deal to acquire a license for this kind of tech.
I suspect the most credible opponents to this startup would be MS bundling this kind of functionality into Outlook, or some other big company exploiting this expensive but already licensed tech.
But what you get here, is first to market, and attack on specific verticals (i.e. social services, and later related services). You can't be scared off by 'what if we have a competitor'. If you don't have a competitor, your market probably doesn't exist. That is what I'd worry about here.
Without relevant domain or speech recognition expertise, it seems that names might be one of the more common areas to get errors while transcribing using such a system, particularly if the system isn't trained for a name and/or it is said by many different voices?
Social workers in the UK have a big problem. Different teams, professions (police, medical, housing), deal with the same people, and often collect all the pieces required to understand and prevent a tragic situation. But they can't put them together.
There is an obsession with interpofessional meetings to tackle this, but every serious case review shows that the knowledge was there but not put together, leading to dead children. My perception is that often knowledge is mentioned, but when discussing a different related case, or isn't formally written down and passed to the right person.
My solution is inspired by the part of the CALO project that dealt with capturing and exploiting knowledge from meetings. There are some great ideas. If you do this, read up on it.
The product records all meetings, using existing meeting room Teleconference hardware. The meetings are put through automatic speech and speaker recognition on a best effort basis to produce a searchable index for that meeting. So you get something like a transcript and can skip to the point where Mr A is mentioned.
Supervisors, who in that field are criminally liable for mistakes by subordinates, already review minutes for meetings. Now they can search for any mention of Mr A, and see in what context he was mentioned and listen to that part of the meeting. Obviously there would be control of who could listen to what meetings, but these people already have big intrusive databases, so I don't think they are too worried.
Ideally, you could then build graphs from there of who is mentioned together, and do some Palantir style stuff to spot associations between people.
But the start is, very much, implementing a meeting transcriber by sowing together a few brands of speech recognition engine, a web interface for administrating and searching meetings, and selling it into social services.