Fact · Languages and accents
Milde and Köhn say that when the recognizer does not know a German compound, the error is counted as at least two errors.
Checked against its source on October 17, 2026
In the source's own words
“each of such errors is counted as at least two errors: a substitution and an insertion error”
Open Source Automatic Speech Recognition for German, Benjamin Milde and Arne Köhn, arXiv (Computation and Language), 2018, Section 3.1, error analysis.
Open the original- What it measures
- The example in the paper is nachzumachen, which the recognizer wrote as nach zu machen. That is three word errors for one spoken word.
- What it does not show
- A 2018 Kaldi system on read German speech. The error mechanism is general, but the counts depend on the scoring method.