Google’s Voice AI Is Now Virtually Impossible To Tell Apart From Human Speech

By Mikelle Leow, 28 Dec 2017

Screenshot via Google

Robots can finally have their say, much to the terror of humankind.

Researchers at Google have developed a text-to-speech tool called ‘Tacotron 2’, which is said to generate speech that’s almost indistinguishable from human talk.

“In an evaluation where we asked human listeners to rate the naturalness of the generated speech, we obtained a score that was comparable to that of professional recordings,” wrote the developers.

‘Tacotron 2’ is able to naturally alter its intonation, stressing on capitalized words like a person would when enunciating more important words.

Fear not—for now. While Google agrees that the AI sounds incredibly “human-like,” it adds that ‘Tacotron 2’ hasn’t been perfected yet.

“[O]ur system has difficulties pronouncing complex words (such as ‘decorum’ and ‘merlot’), and in extreme cases it can even randomly generate strange noises.”

The machine has also only been taught to speak in one female voice. To have the system speak like a male or another female, the researchers would have to retrain it.

Google has released some sample audio clips from the experiment but did not specify which sounds were created by the bot. Do you think you can tell them apart from a regular human being’s?

“That girl did a video about Star Wars lipstick.”