Back to timeline

Benchmark · October 18, 2016

Human parity in speech recognition

Microsoft reported 5.9 percent error on Switchboard, the same rate professional transcribers make.

Why it matters

The task the 1971 ARPA programme set was declared solved for the first time, by a commonly accepted measure.

The bar was not set abstractly: the same material was given to professional transcribers and their error was measured. The qualification matters: parity concerns one corpus of English telephone conversations under laboratory evaluation. Noise, accents and overlapping speakers remained unsolved. The following year the same group brought the figure to 5.1 percent.

Event record

Event date
October 18, 2016
Timeline date
Event date
Verification
Sources gathered automatically · September 17, 2026
Lines
ID
evt-0233

Announcement of 18 October 2016.

Sources

Related events

Records that link to this one