Will a long file lose lines or shift my timings?

No. Each text cue gets a stable ID, and every model response is validated against the expected IDs with unique, non-empty text, so a batch of 40 that returns 35 valid results gets a follow-up request containing only the five unresolved subtitles. The final SRT is rebuilt from the original cue list, and anything still untranslatable stays visible as an empty subtitle and is marked in the QC report.

Can I run the whole thing locally without an OpenRouter key?

Almost. Translation and QC can each target any OpenAI-compatible server such as Ollama, LM Studio, vLLM or llama.cpp by passing a base URL plus the qc-model and qc-base-url options, in which case no OpenRouter key is needed. The exception is Jev itself, since the typesafe models go through OpenRouter's Decisions endpoint. Set LOCAL_API_KEY if your local server requires a key.

How noisy is the Jev review? Will it flag half my file?

It is tuned for precision rather than recall. The reviewer is asked one question per pair and is told to flag missing meaning, unsupported additions, reversed negation, changed or dropped names, numbers, dates or units, and truncated or repeated content, while uncertain or stylistic differences should pass. Sentences that legitimately span cue boundaries are judged with neighbouring cues visible so they are not flagged.

Is this the same as the GeekLink product?

Different scope. This project translates and reviews existing SRT files, with no Python setup beyond a virtualenv and your own model keys. GeekLink is the creator-facing desktop app that also transcribes audio or video, extracts burned-in subtitles, provides an editor and exports finished video, and it is offered as a hosted option.

What does a Jev QC flag actually mean?

A flag is a suggestion to inspect a subtitle, not proof that it is wrong. Jev does not rewrite translations and can miss mistakes; it highlights likely defects such as missing meaning, reversed negation, changed names, numbers, dates or units, truncated content and unsupported additions.

How does the QC review handle meaning that spans cues?

QC runs on ordered batches of up to 40 cues, and before judging one cue the reviewer can inspect as many preceding and following cues in that batch as the sentence or speaker turn requires. Meaning that legitimately moves across cue boundaries is not flagged, and a complete multi-cue sentence should pass in every cue.

What license is the subtitle translator under, and what are the terms?

It is free and open source under the GPL. You provide your own OpenRouter API key, and translation and Jev requests may incur OpenRouter charges; the interface runs locally, but subtitle text is sent to OpenRouter and the model providers you select.

CommunityJev Tools & CLIs

Jev Subtitle Translator

An SRT translator and subtitle QC tool: cue numbers, order and timestamps are preserved locally, missing translations are retried by ID, and Jev flags each source and translation pair that may deserve human review. Web UI or CLI.

See highlights

Open github.com

Stars: 9
Language: Python
License: GPL-3.0
Published by: Community
Author: GeekLinkDev
Last verified: Oct 2026

GeekLinkDev/jev-subtitle-translator

Organization repository on GitHub

View on GitHub

Translate SRT subtitles with structured LLM output and check every translation with Jev.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

English · 简体中文 · 日本語 · 한국어. An open-source SRT translator and subtitle QC tool for LLM-powered workflows. Use it to translate SRT files, preserve every subtitle cue, and automatically check translations with Jev. Topics: SRT translator · subtitle QC · subtitle LLM · OpenRouter · Jev.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Cue numbers, order and timestamps are parsed and held locally; only dialogue text and a stable ID are sent for translation.
Missing or empty results are retried by ID and the SRT is rebuilt from the original cue list, so one failure cannot shift later lines.
Jev answers one question per pair, does this need human review, over ordered batches of up to 40 cues with neighbouring cues visible.
QC runs on OpenRouter's typesafe/jev alias by default, or on any local OpenAI-compatible model, or can be skipped entirely.

Watch out

A Jev flag is a suggestion to inspect a line, not proof that it is wrong, and the README is explicit that Jev does not rewrite translations and can miss mistakes. The review rules are deliberately precision-first, so uncertain or merely stylistic differences pass and a genuinely broken line can slip through. Lines that cannot be translated at all stay in the output as empty subtitles and are marked in the QC report for repair rather than hidden. Translation models must support strict JSON Schema output; small local models miss IDs more often, which is why those lines are retried and then flagged like any other failure, and there is an automatic JSON extraction fallback for models that wrap their object in text. Jev models go through OpenRouter's Decisions endpoint and therefore require OpenRouter even when translation runs locally against Ollama, LM Studio, vLLM or llama.cpp, while any other QC model receives the same guidelines through chat and returns a needs_review verdict per line. Subtitle text and the keys you supply are sent to OpenRouter and to the model providers you choose, and the project is GPL licensed. Scope is existing SRT files only; transcription, burned-in subtitle extraction, editing and video export belong to the separate GeekLink desktop app. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jevclip

Video tool where Jev scores subtitle segments and code assembles a short highlights cut plus a de-watered full cut with timestamped summaries.