AI and Speech Recognition

robonaut-machines-dexterous-humanoid-39644

Time for the robot uprising?

One day, robots just might be thinking and acting on our level. Even though we are not there yet, current strides in the industry have been impressive, to say the least, including Alpha Go, Google’s chiefly known AI system, defeating a Go master in 2017 as well as Luna, the AI, that can give fairly accurate life advice and do math on the level of a mathematician, among other things. All in all, most people in and out of the space seem to be wondering: “when will AI really be just like us, if not more like us?”

But that’s not what I ordered?

This question is more than just difficult to answer, although the process could be said to begin with the latest updates in the field of speech recognition. A key model of the good and the bad in this field is Amazon’s conversational AI system, Alexa. It’s safe to assume that most of us have heard about Alexa ordering things without the user’s consent or simply responding in a wildly incorrect way to commands that it’s given.

Ears to hear and hands to write?

One company that seems to be looking to compete with, as well as possibly even improve upon Alexa is Microsoft. It was reported by several outlets last year that Microsoft set a specific, global record with speech recognition and AI.

All in all, it was made clear that one of Microsoft’s current AI projects can listen to and transcribe from audio as well as a human can. According to Microsoft, they used the key performance indicator of error rate per transcription, which basically just means the percentage of the text that was written down incorrectly. In the experiment done by Microsoft to achieve these results, 2,400 telephone conversations were analyzed but only speakers with what were termed “neutral” accents, were included. Why was the study limited in this way? The reasonable answer seems to be that Microsoft was afraid of tripping up the AI with data that was too far outside of its comfort zone. Even more specifically, this might have been an intentionally controlled variable to keep the error rate as low as possible. If this was the case, this would largely invalidate the study but again, this is just supposition at this point.

Whatever the case was in terms of how the study was carried out, one question still remains. What sort of results would have been obtained if the experiment included a global sample with all kinds of accents? While this question seems to have been left unanswered for now, it is clear that companies like Microsoft are helping to move the AI space forward. In future pieces, we’ll go further into areas like this to highlight some of the biggest innovations in the industry.