Microsoft's new speech recognition system achieves 5.9% word error rate

Microsoft's new speech recognition system achieves 5.9% word error rate

Microsoft's new speech recognition system achieves 5.9% word error rate

According to a research paper published by Microsoft engineers on Monday, the company has reached "human parity" with its new, experimental speech recognition system.

Revealing that the Microsoft's experimental speech recognition system has, this week, achieved the lowest-ever recorded error rate for machine transcription, Microsoft engineers said in their paper that the word error rate (WER) achieved by the system was 5.9 percent
--- the first time the system has recorded a WER below 6 percent. In September, Microsoft had reported a 6.3 percent WER for the system.

The revelation that Microsoft's new speech recognition system has achieved a 5.9 percent WER is noteworthy, especially against the backdrop of the fact that, five years back, the WER achieved by the best speech recognition systems was 20-25 percent.

As a result of the 5.9 percent WER achieved by Microsoft's new speech recognition system, the capability of the system to transcribe conversational speech is now almost at par with humans.

Highlighting the fact that "human parity" has been reached by Microsoft's experimental speech recognition system, Xuedong Huang - the chief speech scientist at Microsoft - said in a statement: "This is an historic achievement."