German speech to text: compounds and scores

By the Potok teamUpdated ~2 min read

German speech to text with Potok is a recommended language. Published large-v3 tests report word error rates from 4.30 to 5.8, from several sources and not from our own measurement. Long compound words and strong dialects are the places where a transcript may need a careful read.

Transcribe a recording

Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.

Open @potok_now_bot in Telegram

Files up to 2 GB · export to TXT, MD, DOCX, PDF, SRT and VTT · speaker labels and timestamps · about 30 languages

Key facts

  • Three sources report German, and the spread of about one and a half points is the honest range.
  • German joins nouns into long compounds, and a word error rate counts a split compound as an error.
  • Regional dialects in speech can differ a lot from standard German, and the figures cover mostly standard speech.
  • Capital letters on nouns are part of the spelling and should appear in the text.

What a transcript looks like

Sample sentence in German
[00:00] Speaker 1: Lassen Sie uns die Besprechung beginnen. Zuerst gehen wir die Ergebnisse der letzten Woche durch, danach planen wir die nächsten Schritte.

Meaning in English: Let's start the meeting. First, we'll review last week's results, then plan the next steps.

Written by us to show the script and the layout of a transcript. It is not the output of the recognition engine.

Published quality

Quality for this language
LanguageVerdictPublished resultNotes
German (Deutsch)Recommendedword error rate 4.3% (large-v3, FLEURS)
Canary-1B-v2 paper, Whale paper, TheWhisper benchmark
Published tests show a low error rate for this language.
Published large-v3 results range from 4.30 to 5.8.

These numbers come from published tests by other people, not from tests by Potok. Your own audio, with its noise, accents and microphone, will differ.

How to do it

  1. Judge the style. Broadcast German is easy; a dialect conversation is harder.
  2. Send the file. Open @potok_now_bot and attach audio, video or a voice message.
  3. Check compounds. Search for places where a compound was split or joined differently from your style.
  4. Large file. Up to 2 GB; use Upload.
  5. Export. TXT, MD, DOCX, PDF, SRT or VTT.
  6. Keep the source audio. A doubtful compound is easiest to settle by replaying the original.
What Potok can do
FeatureDetails
Where it worksTelegram bot @potok_now_bot; large files go through the bot's mini app
File sizeUp to 2 GB
ExportTXT, MD, DOCX, PDF, SRT, VTT
Speakers and timeSpeaker labels and timestamps can be switched on
LanguagesAbout 30 languages with good published results; a few more with warnings
TranslationThe transcript is not translated into another language
DictationA separate Potok desktop app lets you dictate text into any program; it recognises Russian speech

Common mistakes

Common mistakes and how to fix them
MistakeWhat happensFix
Counting a split compound as correctIt may read fine but differ from the reference spelling.Settle on one spelling in your edit.
Using standard figures for a dialect speakerThe tests mostly cover standard speech.Check a clip.
Ignoring noun capitalsThey carry meaning in German.Check the text.
Quoting 4.30 as a promiseIt is the best of a range.Cite the whole range.
Reading the best figure as the typical one4.30 is the low end of a range that reaches 5.8.Cite the whole range.

More

See the full list of languages, the tools and the glossary.

Frequently asked questions

What is the range?

From 4.30 to 5.8 word error rate for large-v3 in published tests by others.

Do dialects work?

The figures mostly cover standard speech, so test a clip from your dialect speaker.

Does it write compounds correctly?

Often, but whether a compound is split or joined is a common difference to check.

Is this our measurement?

No, these figures were published by others and we did not run the tests.

Is the text in German?

Yes, it stays German because the bot does not translate transcripts.

Sources

Each source was opened by hand on the date shown. Pages change, so open the source if a detail matters to you.

  1. Word error rate (Wikipedia), Wikipedia · checked
  2. Whisper speech recognition system (Wikipedia), Wikipedia · checked
  3. Canary-1B-v2 paper, Table 11 (Whisper large-v3 on FLEURS) · checked
  4. Whale paper, Table 5 (Whisper large-v3 on FLEURS) · checked
  5. TheWhisper benchmark (turbo, average over FLEURS, MLS, Common Voice) · checked

Try it on your own recording

Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.

Open @potok_now_bot in Telegram