When translating to arabic it repeats the last lines of the text

When translating to Arabic, whisper repeats the last lines of text in the actual audio file, and it even gives it a timestamp on the actual video, which means it is actually cutting off or leaving out of the transcription the last part of the video.

I tested it with many audios from English to Arabic, and it does the same


I’m using large-v2, i will try with a different one just in case.

Hi @Rrrila

whisper is a speech to text model. It’s meant to transcribe not translate.

Currently the only translations supported are from non-English languages to English, not the other way.

Is it plan to work they other way around (english to non-english or non-english to non-english)?

I’m not OpenAI staff, hence I’m not able to comment on that.

However if you want to translate to non-English languages, you can try using the /edits or completion to translate post transcription.

As stated, im not developer my self so no tsure what do you mean by /edits or completion to tranlate post transcription, if you dont mind pointing me to the right direction so i can investigate myself


Thanks for the info, but unfortunately none of that would help me, as one is for editing the output on specific words or cases, and the other one uses ChatGPT which I’m trying to avoid. I guess they will add support for rest of the languages at some point…

Although I would like to point out this thing does not happen when translating to lets say Spanish, it does it just fine, is on RTL Arabic Language when it goes crazy, so there might be a bug just as I’m saying.

Your assumption is incorrect.

/Edits can edit entire transcriptions to translate text.

Chat completion can do the same for lower cost per token.


I might still misunderstood your approach, but… isn’t that being translated by ChatGPT API? Meaning, there is no way to do so without paying per token, correct?

There is no ChatGPT API.

This is working with the /edits endpoint - Editing is free while in beta.

Here’s pricing for rest of models.