Alexa, tell me a joke
21 Oct 2017
Following last week’s Siri shutout (football result was a double-zip deadlocker in A Field), more interactions with digital assistants today.
We are trying out an Amazon Echo Dot, primarily because changing the station on the “radio” in the kitchen is such a hassle with Sonos. I know, First World problems. If others are interested, integration with Sonos is a new Alexa skill, and if you have existing Sonos devices, it’ll work, but it ain’t spectacular.
Iris tried talking to Alexa. The first thing we noticed: Alexa recognises its trigger word when children speak it (I think Siri is trained to respond to its owner’s voice). The second thing: Iris wanted to try “her” out. So we did, with the same sort of thing we’d tried with Siri.
(audio below is crucial to this blog post)
Iris: “Alexa, tell me a joke.”
Alexa: “Why was the cat so good at chess?”
Iris: “…” (Iris doesn’t know the conventions of humour)
Now, I’m no comic genius, but even I reckon I could deliver this punchline better than Alexa did here. I know it’s a lot to ask of a computerised voice (for now at least), but “tell me a joke” is one of the things that Amazon advertise as something to try out right at the beginning of your Alexa experience. I actually quite enjoyed having to mentally re-parse the line and work out why it was funny. (I had been expecting the Chess-shire Cat, bwahaha)
Actually, I’ve been interested in speech synthesis recently - for work-related reasons - and have been teaching myself about “deep learning” (expect more blogging on this subject soon). If it’s not already possible, I don’t think it’s long before artificial speech is easily able to “correctly” deliver the cat chess punchline. I’m not sure it’ll be much longer before the AI works out that it needs to alter its inflection, without being told.
Some links for interested readers:
- lyrebird.ai - think about ethics while you learn about what’s possible with this tech
- Wavenet generative audio - in this one, the “babbling” is fascinating, but I found the output of the neural net when trained on classical piano, and left to “talk” the most interesting of all