This is interesting to me because it really pulls together a number of disparate technologies to get towards the goal of being "the AI". There's nothing he talks about that's overly complicated on it's own, but integrating them all together is the key.
Once you have the voice to text (ASR) from Nuance (or any of the other companies in this space), then it's a question of properly recognizing the contextual intent. Not a trivial task, but certainly getting easier with technology available. I think the visual models displayed in the demo are fairly telling as to how they handle this. Keywords for a domain (e.g. weather) become useful to determining the intent of the speech, as well as variables used in queries, or the data returned for display to the user.
For example, if I ask a question that mentions temperature, and a time, but not a length of time, it's fairly obvious I'm asking for something in a weather domain rather than a recipe domain (and vice versa).
I'd guess the developer integrations he mentions then become a matter of defining those data points/variables for the different data source so the "AI" can build the application to execute.
It's like an enhanced API integration model. You need to know all of the input/output parameters to integrate with an API. The intent/contextual piece also uses the individual data points for the contextual intent recognition in the voice-to-text area.
The other interesting aspect is actually storing the preferences for the commerce side of things. Airlines do this already for saving preference of aisle vs window seats. They're taking things a step further to remember those types of "qualifying data" for interactions you have so they can be saved across areas (read: API calls).
I suspect there will be a ton of work on behalf of the APIs too handle data in this way too, that's what he says when it will take some time to see the direction things go in. If I opened up my existing airline travel APIs to this today, it's unlikely anyone could correctly interact in a way that would provide all of the information needed to actually book a ticket. So, there will need to be some back-and-forth communication of those missing items. If someone finds a ticket they want to book and says "order it", then viv will need a way for the API to communicate "That's great, but you also need to provide your TSA known traveler ID number". Then, because it's something an API has asked for, Viv will know it's a data point it should save for later.
Let's hope Viv has un-breakable encryption and security with all of those "personal preferences"...
Once you have the voice to text (ASR) from Nuance (or any of the other companies in this space), then it's a question of properly recognizing the contextual intent. Not a trivial task, but certainly getting easier with technology available. I think the visual models displayed in the demo are fairly telling as to how they handle this. Keywords for a domain (e.g. weather) become useful to determining the intent of the speech, as well as variables used in queries, or the data returned for display to the user.
For example, if I ask a question that mentions temperature, and a time, but not a length of time, it's fairly obvious I'm asking for something in a weather domain rather than a recipe domain (and vice versa).
I'd guess the developer integrations he mentions then become a matter of defining those data points/variables for the different data source so the "AI" can build the application to execute.
It's like an enhanced API integration model. You need to know all of the input/output parameters to integrate with an API. The intent/contextual piece also uses the individual data points for the contextual intent recognition in the voice-to-text area.
The other interesting aspect is actually storing the preferences for the commerce side of things. Airlines do this already for saving preference of aisle vs window seats. They're taking things a step further to remember those types of "qualifying data" for interactions you have so they can be saved across areas (read: API calls).
I suspect there will be a ton of work on behalf of the APIs too handle data in this way too, that's what he says when it will take some time to see the direction things go in. If I opened up my existing airline travel APIs to this today, it's unlikely anyone could correctly interact in a way that would provide all of the information needed to actually book a ticket. So, there will need to be some back-and-forth communication of those missing items. If someone finds a ticket they want to book and says "order it", then viv will need a way for the API to communicate "That's great, but you also need to provide your TSA known traveler ID number". Then, because it's something an API has asked for, Viv will know it's a data point it should save for later.
Let's hope Viv has un-breakable encryption and security with all of those "personal preferences"...