Microsoft researchers hit historic speech recognition milestone

The promise of artificial intelligence and the excitement over machine learning and neural nets has captured the imagination of many. As AI systems learn and excel at more complicated tasks, the possibilities grow exponentially. The growth in AI technologies has led to everything from the development of driverless cars to IBM’s Watson helping treat cancer.

But what about AI being able to help you with your daily tasks as a true digital assistant, or other more individual-centric applications of AI like customer service interfaces? This requires machines learning to understand the nuances and complexity of human conversation and language, a very challenging endeavor.

However, Microsoft just announced they reached a huge milestone in conversational speech recognition. In a research paper published yesterday, Microsoft Artificial Intelligence and Research team announced they had achieved an error rate that is on parity with the normal error rates for professional transcriptionists.

In other words, Microsoft’s AI can now understand a conversation just as well as humans who are professional trained and paid to transcribe conversations. Microsoft said:

The researchers reported a word error rate (WER) of 5.9 percent… The 5.9 percent error rate is about equal to that of people who were asked to transcribe the same conversation, and it’s the lowest ever recorded against the industry standard Switchboard speech recognition task.

This marks a huge step forward for Microsoft’s AI efforts and puts a future where we can converse with our computers, rather than using a mouse and keyboard, that much closer. It is also particularly important for Microsoft as the space for digital assistants and IoT devices merge, such as Google Home and Amazon’s Alexa. It will be interesting to see how CEO Satya Nadella looks to commercialize this, as under his leadership Microsoft Research has become very closely linked to the company’s business units.

Share This
Further reading: , , , ,