IBM has broken an industry record with its speech recognition technology, reaching a word error rate of 5.5 percent.
It’s a considerable accomplishment given that the company attained a word error rate of 6.9 percent only half a year ago. More importantly, it has broken the record of a rival, with Microsoft researchers having reached an error rate of 5.9 percent last October.
At the time, the Microsoft researchers claimed they had reached human parity with respect to word recognition. But IBM has since endeavored to take that away from them, teaming with Appen to reassess that industry benchmark. The companies now say that human parity would actually be a word error rate of 5.1 percent.
That, of course, is currently beyond anyone’s reach, but IBM is closer to it than ever. And the new benchmark could further spur a healthy sense of competition as multiple companies race to meet it, given the increasingly important role that voice and speech recognition play in consumer electronics and the emerging Internet of Things.