How do I get whisper to not-translate

[ArthurPeabody]

I made a transcription of an audio mostly in English but about 90 seconds of German. whisper translated it into English. I wanted the German. I didn't use the -tr switch.

[yudin-s]

This is probably not the -tr flag. In whisper.cpp, -tr / --translate explicitly asks Whisper to translate the source language to English, but the CLI also has a separate language setting. The current CLI help shows -l LANG, --language LANG and its default is en.

So if you ran the default command, or used an English-only model such as ggml-base.en.bin, the decoder is being pushed toward English output. A German section in a mostly English file can then come out as English even though you did not pass -tr.

For German text to stay German, use a multilingual model, not an .en model, and do not force the whole file to English. For example:

./build/bin/whisper-cli -m models/ggml-base.bin -l auto -f audio.wav

If the audio is mostly English with only one German section, auto-detection may still choose English for the whole file. In that case the most reliable workaround is to split the 90-second German part and run that segment separately with:

./build/bin/whisper-cli -m models/ggml-base.bin -l de -f german-part.wav

Then run the English sections with -l en or -l auto and stitch the transcript/SRT back together. Whisper is not very good at switching the forced language mid-file when one language dominates the context.

Short version: use a multilingual model and -l auto/-l de; avoid .en models and avoid the default forced en language for the German section.

If this explains the English output, please mark it as answered so others know to check the model and --language setting, not only --translate.

[ArthurPeabody]

Of course I used a non-English model. I needed -l de . I could point to the target section with -ot & -d Thanks.