[in9]

From what is worth, I'm a data scientist who has worked on ads in the past. I've always been under the impression that massive scale audio processing for ad targetting was way too expensive when compared to the signal we can extract from the audio data and as compared to the final ad conversion. I mean, at least from the prices of having an instance processing streams/bacthes of audio data just to sort out brands/products/services that were mentioned in those audio streams.

In my understanding, showing ads for what people that are connected to you searched/bought/interacted, as in a simple network analysis, would be much cheaper and would give you very similar results.

[phkahler]

It's not expensive at all when run on the edge device. Alexa can identify its name locally. You dont need dictation level accuracy to pick out advertising words, and if you miss some it's totally OK.

[tensor]

This paper is studying how requests to the smart speakers are used. Smart speakers translate requests into text, and I'm assuming the ads are just keyed off the resulting requests just like they would be if you put them into the search system manually.

[wutbrodo]

Given your framing as "massive scale audio processing" to pick signals out of audio streams, I think you may have misread (or not read...) the paper.

It's not claiming Alexa is listening in while inactive. It's claiming that it's inferring user characteristics (age, health, etc) from the audio of voice commands, instead of just the post-transcription text.