Microsoft reported on another set of improvements in their speech recognition system. The upgrade sets a new personal record; now processing of voice data reveals the lowest count of misrecognised words which is 5.1%. Not so long ago, October last year, the mark was on 5.9% which at the time was compared with the accuracy of a professional human typist.
During the both experiments they transcribed recordings from Switchboard corpus. Switchboard includes about 2,4 thousands pre-recorded phone conversations. This system is used for testing of voice recognition systems from the beginning of the 90s.
The latter experiment was conducted in order to determine precision levels of the improved system. It was compared to the precision level of a professional typist, who had to outline the context of conversation after hearing the material several times, at the same time working with their colleagues.
In comparison with the previous year, the occurrence of mistakes was reduced by almost 12%. The new achievement was made possible due to improvements in acoustic and linguistic systems based on neural networks, a computer system modelled on the human brain and nervous system.
The Microsoft speech recognition system is used in such services as Cortana, Presentation Translator, and Microsoft Cognitive services.