Preprint
Open Source Automatic Speech Recognition for German
Benjamin Milde and Arne Köhn, arXiv (Computation and Language), 2018
Read the originalRead on October 17, 2026
- What it is
- Milde and Köhn trained German acoustic models with the Kaldi toolkit on two openly licensed German datasets and analyzed where the recognizer makes errors. Their error analysis found that compounds were the largest problem.
- How much weight it carries
- A peer-reviewed-style academic preprint from a university group. It describes a 2018 Kaldi system on read speech, not today's engines, so the error types are more reliable than the exact error rates.
Facts taken from this source
- Milde and Köhn say that when the recognizer does not know a German compound, the error is counted as at least two errors.
“each of such errors is counted as at least two errors: a substitution and an insertion error” (Section 3.1, error analysis)
- Milde and Köhn say that compounds unknown to the language model are quite frequent in German, because German is a very productive language.
“As German is a very productive language, compounds unknown to the language model are quite frequent” (Section 3.1, error analysis)