Why doesn't Alexa know that homonyms aren't homophones?
By @edent
· AI alexa · 3 comments · 250 words · read ~250 times.
As we head unto an AI dominated future, the Turing test will probably become less like a Voight-Kampff test and more like a warzone Shibboleth.
Yesterday, I asked the Alexa to set a timer.
"What do you want to name your timer?" She It asked. "Bow," I replied. "Bow timer set," it said.
Except… that isn't quite right. I wanted a timer for my bāo buns (包). That's pronounced /baʊ/ - as in to bow one's head.
So Alexa translated my speech to text and stored it as "bow".
When it came to read back the word, it pronounced it as /bō/ - as in a bow and arrow.
And there was much confusion.
I've ranted before about Alexa's complete lack of common sense and its inability to handle any form of nuance. The ability to store the original pronunciation of a homonym isn't even AI - it's barely even a database!
A common trope in World War 2 movies is that enemy soldiers are unable to pronounce normal English words like "Loughborough", or "Worcestershire sauce", or "Mr. Cholmondley-Warner".
Will it be so with AI?
Perhaps we will find ourselves being quizzed by an interlocutor that demand we prove our humanity by assessing whether words rhyme?
An arms race, to be sure. But one the machines show no signs of winning. Yet.
ISLA SALISBURY says:
Interesting I found a new mis pronunciation of my name ISLA pronounced I-la. I Often get human mispronunciations of Iss-la not realising it’s a silent ‘S’. Tried out the spotify AI -DJ who decided I’m ill -i-ya. Aside from that it decided Glen Miller, followed by Imagine Dragons then then Sex Pistols was a good mix!
Alex Gibson says:
Similarly, Google Maps, which I otherwise deeply love, is into its second decade of pronouncing the name of my nearest town, Reading, like the verb 'reading'. We can't reasonably expect it to get every weird and wonderful place name in the UK and around the world right. However, would it kill Google to put an option, buried in a three-dot menu, to submit crowdsourced pronunciations in phonetic spelling? rɛdɪŋ - Set some kind of sensible threshold for corroboration/human moderation to weed out nonsense, and over time, the issue would be solved. Not doing so is just lazy. Having left a few reviews for local businesses I'm spammed by them soliciting reviews for other things they've noticed I've driven past - but I've no way to correct the name of the town and its local roads. This isn't an AI job as you say, it's just improving the fitness for purpose of the current algorithms and database.
@Edent I once asked our Google home assistant to "play John Cage's Four Minutes thirty three seconds". After a pause, it replied "I can't find any piece called four feet thirty three inches..."
More comments on Mastodon.