How to transcribe a long German interview accurately

The biggest accuracy gains live before the recording starts. A USB headset mic with a foam pop filter, positioned 20-30 cm from the interviewee's mouth, will give you a transcript several percentage points better than the same speaker into a laptop mic from the other side of a kitchen. Hard surfaces (glass, tile, bare walls) bounce audio; soft surfaces (curtains, carpet, a bookshelf full of books) absorb it. Move to a soft-surface room if you have the choice.

Vosk Small is fast, runs in seconds, and is good enough for note-taking and topic indexing. Vosk Large is roughly 5x slower but meaningfully more accurate on proper nouns, technical vocabulary, and slightly noisy audio. For an interview you'll quote, pick large.

Run through the transcript once with Cmd+F on every proper noun the interviewee said. Names, organisations, place names, acronyms. These are 70 % of the corrections you'll make. Doing them as one pass before the prose pass is faster than fixing inline.

Read the transcript end-to-end while listening to the audio at 1.5x. Stop only at the lines that don't make sense. This is the 'speaker said X but the model heard Y' pass; 15-20 minutes of real work for an hour of interview. Sprachmemo keeps your edits separate from the original model output, so you can re-transcribe with a different engine later without losing your corrections.

Markdown export gives you the title, your edited body, and every clip's transcript inlined under H2 headings. Drop it into your CMS, your team wiki, or your private archive - whatever respects Markdown.