genie-toolkit

genie-toolkit copied to clipboard

AST manipulation methods are tricky and full of corner cases. Testing by generating sentences is not exhaustive, unless we run at high depth and high pruning size, which takes too...

gcampax

It seems to be invoking the Python script for every single sentence to autoparaphrase, which causes the model to load from disk and initialize, and doesn't take advantage of sentence...

gcampax

Device discovery was removed from thingengine-core when it was merged here, because the dialogue loop cannot be represented well in terms of ThingTalk code.

gcampax

Currently the agent will pick one, two or three results from the top of the list (depending on the parameter priority and some randomness in pruning). It would be nice...

gcampax

Support richer output

1

Our system should allow developers to specify non-textual content in the replies from the agent. For speech, we should support sound effects, emphasis, prosody. We might want to target [Speech...

gcampax

`genie` `` ... ?

gcampax

quoting for mturk inputs

2

Entity strings in the sentences are not quoted when generating mturk inputs.

sileix

We need a better wakeword implementation

1

That's it. That's the bug. Nothing will work until we have a decent wakeword. The current one is just bad.

gcampax

On a Raspberry PI, echo cancellation is so CPU intensive that PulseAudio can hit buffer underruns, which cause audible cracks in playback.

gcampax

This is the close cousin of https://github.com/stanford-oval/thingtalk/issues/231 Of course, at first turn the work is basically done, because it is already working in single-sentence datasets, but we should be able...

gcampax