Another milestone in Microsoft speech recognition, error rate is the lowest in the industry

Recently, Microsoftâ€™s chief speech scientist Huang Xuedong stated in the paper that they achieved a technical breakthrough of word error rate (WER) as low as 6.3% in the recent industry-standard Switchboard speech recognition benchmark test, which is 6.6% faster than IBMâ€™s last week. The WER dropped by 0.3%, reaching the lowest error rate in the speech recognition field.

Huang Xuedong mentioned in the paper:

Our best single system achieved a 6.9% word error rate on the NIST 2000 Switchboard set. We believe this is the best performance that a single system can currently achieve. On Switchboard test data, the combination of multiple speech models will further reduce the word error rate to 6.3%.

Microsoft said that the principle of biological operation of the human brain inspired Microsoft to develop a deep neural network to achieve this breakthrough. In addition, they also mentioned the impact of deep learning algorithms of the Computing Network Toolkit (CNTK) on this technological breakthrough, and the contribution of GPU clusters to this advancement. Microsoft pointed out that Cortana can absorb more than ten times of speech data in the same time.

Microsoft's speech recognition technology has become more mature. Under the premise of continuously breaking through the basic technology of speech recognition, it continues to enhance various application services based on speech recognition technology, including Cortana and Skype translation, making humans and machines more natural Interaction.