whisper.cpp icon indicating copy to clipboard operation
whisper.cpp copied to clipboard

--translate is listed as false, but is applied anyway

Open tkzv opened this issue 10 months ago • 2 comments

./main -h says that translation to English is off by default:

-tr,       --translate         [false  ] translate from source language to english
...
-l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)

But when I transcribe a file with non-English speech, by default all of it is automatically translated. Specifying -l auto sometimes disables translation. Setting the language explicitly translates everything to the forced language. Often poorly.

How do I disable translating the wrong language parts?

whisper.cpp version: 1.7.3. Model: ggml-large-v3-turbo-q8_0.bin. OS: Gentoo Linux.

tkzv avatar Dec 29 '24 01:12 tkzv

I was encountering this same issue with whisper-stream myself. Sometimes the text was translated.

I tested it with Arabic. Sometimes the text was translated. Othertimes it would say "(Arabic)". Othertimes it was transliterated. Its been pretty inconsistent.

Aseeef avatar Jan 14 '25 22:01 Aseeef

Had the same issue. It can be fixed by specifying your language manually with -l, and it will dump the original output. I didn't have the issue of the text being translated from english, though.

I'm using v1.7.4 with large-v3.

ocaballeror avatar Jan 31 '25 11:01 ocaballeror

Hey! Same problem here with this line:

./build/bin/whisper-cli -oj -mc 0 -l fr -tr false -m ./models/ggml-large-v2.bin -f ${input} -of ${output}

Always get the text translate to english automatically. I tried with different models (V2, V3, base, etc.).

rdhox avatar May 19 '25 12:05 rdhox