W3C

– DRAFT –
Voice Interaction

10 January 2024

Attendees

Present
debbie, dirk, gerard
Regrets
-
Chair
debbie
Scribe
ddahl

Meeting minutes

dirk's implementation

dirk: no client yet
… ipa registry
… ipa provider is chatGPT
… using client input, sends text to dialog layer
… the concepts are working
… creates provider selection service

implementation is C++

dirk: only uses "multimodal input" parameter
… this is a container of various multimodal inputs
… returns client response
… session id, request id, multimodal output
… the ipa service must know about the provider selection service
… provider selection service knows about the provider registry
… provide registry is hard-coded, knows about providers
… chatGPT is one provider
… uses provider selection strategy, which hasn't been implemented yet, just uses the first one
… eventually will need to use keyboard and display in client
… need to add more logic about how to select IPA

debbie: seems like picking the IPA is the job of the provider selection service
… not the registry
… location of chatGPT is part of the adapter, not the registry

debbie: could use a similar pattern for Amazon Lex, for example

dirk: provider selection strategy should know about supported modalities
… could try to match input with output
… adapter should mention what it supports
… in terms of modalities
… we never discussed error situations
… for example, ChatGPT gave an error for incorrect credentials

dirk: should return exceptions if the input is bad
… rather than error codes
… we need to decide this strategy soon
… this is in the repository, where to put the documentation

debbie: should be in the "source code folder

dirk: what we've defined so far seems to be working

debbie: next priority is provider selection service and registry need to be fleshed out
… maybe the client isn't that critical

dirk: would need that if we don't want to recompile every time

dirk: need to check license of source code

debbie: check to see if w3c has a preference
… could use Apache 2.0 if the W3C doesn't have a preference

dirk: then elaborate on provider selection strategy
… decide on return values vs. exceptions

debbie: providers should advertise what modalities they support

dirk: this would be part of the provider selection strategy
… highest priority is to document in the README
… invite others to support

dirk: still planning to look at dialog events

debbie: should also look at conversation envelope spec, which will be published in the next few days

dirk: didn't run into any major problems with implementation

dirk: should create issues for next steps
… any questions from gerard? security?

gerard: implementation of small LLM's on smartphone
… European project called "Evita", coaching older people
… want to avoid cloud access

dirk: would it make sense to add this as a wrapper around your project?

gerard: not sure

Minutes manually created (not a transcript), formatted by scribe.perl version 221 (Fri Jul 21 14:01:30 2023 UTC).

Diagnostics

All speakers: debbie, dirk, gerard

Active on IRC: ddahl