translate — Translation, Without the Timeline Drama

Got an intermediate file in a language you don’t want? translate rewrites the text only — the timeline, the word details, everything spatial stays untouched. It’s text alignment, not a remix.

Centurion translate <CENTURION_FILE> -t <LANG> [options]

Options

• <CENTURION_FILE> — input Centurion intermediate file (.centurion.json)
• -t, --target-language <LANG> — target language code, e.g. zh, en, ja (required)
• -o, --output <OUTPUT_FILE> — output intermediate path (defaults to <input>.translated.centurion.json)
• --source-language <LANG> — source language (auto-detected when omitted)
• -s, --strategy <STRATEGY> — translation strategy, llm (default)
• --model <MODEL> — LLM model (provider default: OpenAI gpt-4o-mini, DeepSeek deepseek-chat, Ollama llama3.1…)
• --api-key <KEY> — API key for the chosen provider; omit it to fall back to local Ollama
• --llm-provider <PROVIDER> — pick the API: openai · deepseek · moonshot · zhipu · openrouter · groq · siliconflow · dashscope · ark · azure · ollama (default auto)
• --llm-base-url <URL> — override the provider’s default endpoint (e.g. a self-hosted gateway)
• --glossary <FILE> — glossary JSON: { "source": "target" } or [{ "source":…, "target":… }]
• --target-script <FILE> — target-language script (one line per subtitle)
• -b, --bilingual — output bilingual subtitles (source on top, translation below)
• -k, --karaoke — word-level \K timestamps for translations

The magic of target-script alignment

Hand it an official target-language script (subtitles count matching) and translate performs pure 1:1 text alignment — line N of the script becomes line N of the subtitles, no LLM involved. Official wording, guaranteed. If the counts don’t match, it warns you and translates with the script as a wording reference instead.

Translated karaoke timestamps ⏱

No word-level timing exists for a translation — so Centurion builds one:

• Time interpolation: the sentence’s [start, end] is redistributed across the translated words
• Long-syllable bias: longer words (more syllables / more characters) get proportionally more time — no more “a” and “extraordinary” fighting over the same millisecond
• A lead-in pause (\K) opens each line, matching the classic karaoke rhythm

Examples

# Convert first, then translate
Centurion convert subtitles.srt
Centurion translate subtitles.centurion.json -t zh --api-key sk-xxx --model gpt-4o-mini

# DeepSeek (cheap & fast)
Centurion translate subtitles.centurion.json -t zh --llm-provider deepseek --api-key sk-xxx

# Moonshot / Kimi
Centurion translate subs.centurion.json -t zh --llm-provider moonshot --api-key sk-xxx --model moonshot-v1-8k

# Local Ollama (no key needed)
Centurion translate subtitles.centurion.json -t ja

# Glossary + official script + bilingual + karaoke
Centurion translate subs.centurion.json -t zh --glossary terms.json --target-script official.txt -b -k

# Build the final bilingual karaoke ASS
Centurion build subs.translated.centurion.json

LLM providers — Centurion speaks the OpenAI-compatible dialect everyone uses. Set --llm-provider (or just --llm-base-url) and the right defaults are applied automatically; auto infers the provider from the endpoint host, or falls back to OpenAI when an API key is present and to local Ollama otherwise. Provider names are forgiving: ds, kimi, glm, silicon, aliyun, volcano… all work. The same options power LLM sentence splitting in asr and ocr (-s llm --llm-provider …).

Bilingual layout: source line rides on top (Default style), translation hugs the bottom (Sub style) — a layout borrowed from classic dual-language fansubs.

Every run saves a <output>.translated.centurion.json — full config, per-sentence translations, glossary & script load state, and any warnings. Perfect for debugging “why did that line come out like that”. Chain it into dub or build directly.