You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Korean. Translations can now go into Korean.
Korean tracks follow the Netflix Korean style guide: 16 characters per line, at most 12 per
second, with Latin letters, digits, spaces and punctuation counting half.
The prompt asks for the polite speech level (해요체) unless the context calls for another,
names written in Hangul, and standard Korean spacing.
Several target languages at once. In Çevir, pick target languages as chips, for example
English and Korean. Each language becomes its own track; the video context is read only once.
Japanese is new in the list too.
Korean fonts.
The style panel warns when a Korean track's font has no Hangul and suggests one that is
installed (Noto Sans KR, Malgun Gothic, Apple SD Gothic Neo, …).
In bilingual mode the second track has its own font. A Korean second track gets a Hangul font
automatically, so the preview and the burned-in video do not show boxes.
Local engine (Ollama). Correction and translation can run on a local model: free, offline and
without rate limits.
Tercihler has the default engine, the Ollama address and the list of installed models.
Düzelt and Çevir can switch engine for each run.
The welcome screen shows whether Ollama is running.
Free-tier mode (Tercihler): one request at a time, batches of at most 15 lines, and a pace
that respects the per-minute limit.
Altyazı → Başarısız satırları yeniden dene sends again only the lines the model did not
finish, so lines that already finished are not paid for twice.
Fixed
A correction or translation no longer stops and loses its finished lines when the model stops
answering.
A batch with no answer is split in half and retried. A line that still fails is marked and the
run goes on.
After three failures in a row the remaining lines are marked without calling the model again.
Rate limits are waited out for as long as the API asks.