translate-transcript

Translate transcript

Translate a transcript file end-to-end while preserving every structural element. Source file is never modified — output is a new file with a language-suffix.

Inputs

• A transcript file (any stage: .raw.md, .cleaned.md, .structured.md, etc.)
• Target language (ISO 639-1 code preferred — en, he, es, fr, de, ar, etc.; full names also accepted)
• Source language is auto-detected unless overridden

Output convention

<stem>.<source-lang>-to-<target-lang>.md

Examples:

• recording.structured.md → recording.structured.he-to-en.md
• interview.cleaned.md → interview.cleaned.en-to-es.md

If the source filename already encodes a stage (.cleaned.md), keep that stage suffix; the language suffix goes after.

Backend choice

Backend	When
Claude (current session)	Default — high quality, preserves nuance, no extra config
Gemini via OpenRouter	When transcript is very long (cheaper at scale) and OPENROUTER_API_KEY is set
External MT API (DeepL, Google Translate)	Only on explicit request — generally lower quality for nuanced/idiomatic content

Preserve, don't translate

These elements pass through unchanged:

• Markdown structure: headings, lists, blockquotes, code blocks, tables
• Timestamps: [00:12:34]
• Speaker labels: **Alice:**
• Inline code, URLs, file paths, technical identifiers
• Proper nouns (names, brands, places) — translate only if there's an established target-language form (e.g. "Jerusalem" / "ירושלים")

The italic source-attribution header stays in the source language and gains a translation note:

_Topic: …_
_Translated 2026-04-28: he → en, Claude_

Translation guidance

• Register: match the source's register. Casual voice memo → casual translation. Formal meeting transcript → formal.
• Idiom: translate idioms by meaning, not literally. Hebrew "על הפנים" → "terrible" / "awful", not "on the face".
• Hedging and filler: if the input is a .cleaned.md (filler already removed), don't add filler back.
• Cultural references: keep the reference, but add a brief parenthetical gloss only if the target audience genuinely won't recognise it.
• Code-switched content: if the speaker switched languages mid-sentence (common in Hebrew-English mixed memos), translate the whole result into the target language rather than preserving the switch.

Long inputs

For transcripts > ~30 minutes / >10k words:

• Translate in chunks of ~2k words, overlapping by one paragraph for continuity
• Re-stitch chunks, removing the overlap from the second chunk
• Pass a glossary of recurring proper nouns / technical terms to each chunk for consistency

Output report

Source:       transcripts/recording.structured.md (he, 4,820 words)
Target:       transcripts/recording.structured.he-to-en.md (en, 4,610 words)
Backend:      Claude
Glossary:     12 proper nouns held consistent

What not to do

• Don't summarise. Translation preserves length and content.
• Don't "improve" the source. If the speaker rambled, the translation rambles.
• Don't drop or merge speaker turns.
• Don't translate file paths, command names, or code — only natural-language content.