Incorrect timings
jrkoiter opened this issue · 2 comments
jrkoiter commented
Cool program! When translating this subtitle from English to Dutch:
cli/translator.mjs --stream --temperature 0 --from "English" --to "Dutch" --file ~/Downloads/subtitle.srt
The resulting timings are off:
Cerlancism commented
Thanks for trying out the program. I am aware of this problem, this problem resides in ChatGPT itself.
There are several workarounds I thought of:
- Cross-check manually. I use Google Sheets, paste in the original
.srt
(copied from Aegisub, which writes CSV to the clipboard), paste in the outputprogress
.csv
, and compare it with another translation service, such as the =GOOGLETRANSLATE formula - Use smaller batch sizes, or even
--batch-sizes [1]
- Provide a more specific system instruction, such as the one found here, which burns more tokens. However, the result is still not guaranteed.
https://github.com/gnehs/subtitle-translator/blob/52ee34917dad5aa0a8eb7dd89b9b971b8438035c/index.js#L34
Cerlancism commented
This issue could be improved with the newly supported structured mode
Experimental feature:
Try in CLI --experimental-structured-mode