Ever get frustrated with your phone's personal assitant not being able to understand you? Or have you received a transcribed voicemail that made no sense? Those days are numbered thanks to innovations at Microsoft - their AI's voice recognititon engine error rate has now reached an all time low of 5.1%; comparable to the error rate of humans' comprehension of speech. Microsoft credits "neural-net based acoustic and language models" for this improvement, as well as giving its AI the ability to analyze context when deciphering speech. For example, the phrase "let's leave it red" could be interpreted as "let's leave it read", which could prove to be a confusing transcription for recipients. However, if the AI has the opportunity to consider to the entire block of speech and hears the word "logo", "colors" and the like, it can safely assume that the conversation is about a design decision and "red" is the correct word. Or, if the speech talks about messages or replying, it can understand that "read" is correct. Microsoft technical fellow Xuedong Huang, who reported the error rate success, says that this has been a Microsoft research goal for 25 years.
Though a technological breakthrough, Microsoft's engine still faces challeneges in understanding speech in noisy environments, accents, and colloquial phrases. Plus, this success hasn't spread to other languages yet either, as each language handles individual elements differently and there may be limited training models available for less-spoken languages. However, Huang adds "Moreover, we have much work to do in teaching computers not just to transcribe the words spoken, but also to understand their meaning and intent. Moving from recognizing to understanding speech is the next major frontier for speech technology."
It remains to be seen just how influential AI understanding of speech will be, but it's poised to be a major technological development when it happens.
This article was based on an August 21, 2017 Business Insider article by Rob Price. Read it here.