subtitle-studio

Translating subtitles into multiple languages sounds straightforward until you actually do it: parse SRT or VTT, call an LLM sentence by sentence, merge the bilingual output, validate alignment, and repeat across an entire directory — all while fighting encoding quirks, malformed timecodes, empty translations, and timeline overlaps. subtitle-studio packages this whole pipeline into a single zero-runtime-dependency workflow. Its subtitle parsers are built from a ground-up tolerant state machine that copes with missing sequence numbers, CRLF line endings, UTF-8 / UTF-16 BOMs, and malformed timecode lines, always producing clean UTF-8 output with millisecond-accurate timing.

The translation layer speaks any OpenAI-compatible HTTP endpoint (DeepSeek by default) or the native dsh ctx.llm seam, batching cues by a character budget, aligning results through JSON payloads keyed by cue index, retrying with bounded attempts, and injecting a JSON glossary as mandatory terminology. Bilingual merging offers stacked or interleaved layouts with the timeline copied verbatim from the source so it cannot drift. Alignment validation checks count mismatches, missing or extra cues, empty translations, overlong cues, and timeline overlaps. Directory batch mode adds a bounded concurrency pool, an atomic checkpoint for pause-and-resume, and per-model token cost estimation, with the output tree mirroring the input so same-named files never collide.

Whether you are a creator batch-translating course or film subtitles, or a developer integrating subtitle translation into a dsh workflow, subtitle-studio exposes the same engine through five dsh tools (sub_parse, sub_translate, sub_merge, sub_export, sub_glossary) and a full-featured CLI — turning subtitle translation from a custom script exercise into a configuration task.