Tech Oxymoron #3: Speech2Face

Apparently the safest way to be racist, sexist, ageist or grossly prejudiced is to be an AI that is born in MIT and hyped as ground breaking by its creators. Imagine if you closed your eyes and heard someone speak and decided how they looked. Perhaps black, perhaps chinese, perhaps illegal alien, perhaps dangerous?

MIT’s Speech2Face is being hyped as nearly there.

Here are their claims for real world applications:

• Cartoon representation of person in conference call where the person does not want their actual face seen :) Hmm, but a sketchy likeness is ok?
• Faces for Alexa and Siri. Why not just assign our best friends’ faces?
• Law enforcement by creating a portrait of suspect if the only evidence is a voice recording. How many ransom callers use their own voices anyway?

Here are their admitted short falls:

And yet they claim their AI is uncannily accurate :) And that accuracy looks like this: