The past months have witnessed breakthrough announcements from Microsoft, IBM, and Google, all hitting new marks in speech recognition accuracy; they claim that the error rate has reached 5.1 percent — the word error rate of humans.
Still, that doesn’t seem particularly accurate — the last time I spoke to a machine, I didn’t get the feeling that the recognition was nearly that good. Let’s review the next generation of speech technologies, how they are enabling new analytics, and the growing role of these insights for businesses to see whether the companies’ claims are accurate.
In early 2000, speech recognition reached 80 percent accuracy. In the enterprise space, it triggered adoption for Interactive Voice Response (IVR), which was initially implemented to remove the complications involved in customer service problems.
But speech applications were very dependent on vocabularies and languages. They required a sophisticated set-up by highly specialized system integrators, and each major language had its speech recognition startup. It was only in 2005, when Nuance snapped up 15 companies, that the space was consolidated.
See also: Amazon makes it cheaper for developers to use Alexa voice
Aside from speech usage for IVR, a second use case emerged for Quality Management (QM). Customer service organizations use quality management applications to listen to call center calls and rate them. The process used to be tedious, limited to a small sample of calls, and like looking for a needle in a haystack. With speech recognition, it became possible to automate parts of the process. Workforce optimization leaders NICE and Verint developed or bought ways into speech analytics, followed by Contact Center Infrastructure players like Avaya or Genesys.
These developments remained limited. IVR speech enablement has failed to transform the customer experience, and voice self-service experiences continue to be rated poorly. Speech for QM is often confined to compliance or script adherence verifications. At the beginning of the 2010s, it seemed that speech technology had stalled.
While speech for customer service was developing, Amazon, Apple, Google, IBM, and Microsoft continued investing in research and development of speech technologies, driven by the vision it would eventually become critical for user interaction with machines.
Apple broke into the market with the introduction of Siri, which used machine learning to transform speech recognition. Artificial Intelligence removed many of speech technology’s complications and intricacies, as well as the need to re-engineer the stack for new languages or new vocabulary sets.
Today, most digital disrupters, most notably China “Big Three,” Alibaba, Baidu, and Tencent, are building their speech stack. Because generic machine learning engines can be used, barriers to entry have been lowered dramatically. Open source options, like CMUSphinx, HTK, Julius, Kaldi, and Simon, are also widely available.
The AI breakthrough has paved the way for new entrants. Companies like iFLYTEKor Speechmatics are aggressively addressing the issues of usability, accuracy, and deployability, in particular beyond the dominant languages.
For customer service, the battle is now shifting to the other half of the equation, Natural Language Processing (NLP) and Natural Language Understanding (NLU).
