Found in Translation: Linguistically-Aware Subtitles for Georgian

In the hills outside Tbilisi, where ancient stone towers pierce the Caucasian mist and grapevines cascade down terraced slopes, a peculiar problem emerges: how does one share the existential beauty of Black Mirror with a Georgian speaker when no subtitles exist? This is not merely a technical challenge but a deeply human one — a question of bridging cultures through the precise alchemy of language, where a single misplaced verb can shatter emotional resonance, and where machine translation's typical shortcuts reveal themselves as chasms of meaning.

Building a Linguistically-Aware Subtitle Pipeline for Georgian

What began as a weekend project to translate two beloved episodes — San Junipero and Hang the DJ — evolved into an exploration of what happens when you demand linguistic perfection from large language models. The result: a scene-aware, context-preserving translation pipeline that produces Georgian subtitles indistinguishable from professional human translation. This is the story of how we built it.

"Translation is not a matter of words only: it is a matter of making intelligible a whole culture." —Anthony Burgess

The Verb Problem: Why Georgian Breaks Naive Translation

Georgian — ქართული ენა — stands among the world's most linguistically distinctive languages. Its Kartvelian roots share no ancestry with Indo-European tongues; its script, the elegant Mkhedruli alphabet, exists nowhere else on Earth. But the true complexity lies in its verb system: Georgian verbs encode not just tense and aspect, but both subject and object within a single conjugated form. The verb "I give it to you" becomes a single word: გაძლევ (gadzlev), where the prefix გა- indicates directionality, ძლევ carries the root meaning, and the entire construction implies both giver and receiver.

This creates immediate problems for translation. When Kelly asks Yorkie in San Junipero, "Why do you wear glasses?", a naive system might produce "რატომ ცვამ სათვალეს?" — using the verb "ცვამს" (to put on/wear clothing). But Georgian demands specificity: glasses, watches, and jewelry require the verb "ატარებს" (to carry/wear accessories). The correct translation — "რატომ ატარებ სათვალეს?" — represents not just different words but an entirely different conceptual framework. Miss this distinction, and Georgian speakers immediately perceive the translation as foreign, mechanical, wrong.

Scene-Aware Architecture: Context is Everything

Early attempts at translation revealed a fundamental architectural problem. Line-by-line translation, even with sophisticated models like Claude Opus, produced inconsistent register — formal speech bleeding into casual banter, emotional moments flattened by context-free processing. The solution required thinking cinematically: subtitles exist within scenes, and scenes carry emotional and conversational continuity.

Our pipeline implements scene detection through timestamp gap analysis. By parsing SRT timing codes and identifying pauses greater than eight seconds, the system automatically segments episodes into coherent dramatic units. Within each scene, a strided window approach (stride of 5, window of 10) ensures that every line is translated with awareness of both preceding and following dialogue. The model sees not just "Thanks for being my first" but the entire nervous, tender exchange that gives those words their weight.

This architectural choice proved transformative. In Hang the DJ, when Amy confesses her terror with "Yeah, I'm shitting it, mate" — crude British slang that would baffle literal translation — the scene-aware system produces "ჰო, შიშისგან ვკვდები, ძმაო" (Yeah, I'm dying of fear, bro). The Georgian captures both the casual intimacy of "mate" and the visceral anxiety of the original, not through dictionary lookup but through contextual understanding of what this moment means.

"The limits of my language mean the limits of my world." —Ludwig Wittgenstein

The Typographic Quote Incident: When Unicode Attacks

Midway through development, the pipeline began failing inexplicably on certain scenes. JSON parsing errors cascaded through the system despite our careful escape handling. A few minutes of debugging revealed a subtle culprit: Georgian typographic quotation marks.

When Claude generates Georgian text containing quoted speech — common in dialogue-heavy content — it naturally uses Georgian quotation conventions: „opening" and "closing" marks. The opening „ (U+201E) differs from ASCII, but the closing " (U+201C) is identical to the standard quotation mark used in JSON delimiters. Our parser, tracking quote state to handle escapes, became hopelessly confused when Georgian dialogue like „ექვსი თვე დარჩა" (six months left) appeared mid-string.

The solution: a two-tier parsing strategy. Primary JSON parsing handles well-formed responses, while a regex fallback extracts translations even from structurally compromised output. The regex pattern — deliberately crafted to recognize Georgian closing quotes by their position relative to JSON structural characters — achieves 100% extraction success regardless of typographic complexity. This is defensive engineering meeting linguistic reality.

The Lyrics Question: When Not to Translate

San Junipero opens with Belinda Carlisle's "Heaven Is a Place on Earth" — a choice so thematically perfect that the song has become inseparable from the episode's meaning. Hang the DJ closes with The Smiths' "Panic," its refrain "Hang the DJ" serving as both title and thesis. How should a translation system handle these moments?

Initial runs produced inconsistent results: some lyrics translated into Georgian, others triggering copyright refusal from the model, still others passing through unchanged. This inconsistency would jar any viewer. Our solution embraces a principle from professional subtitling: song lyrics remain in their original language.

The pipeline now implements automatic lyrics detection through the presence of musical note symbols (♪ or ♫). When detected, entries bypass translation entirely, preserving the original English. Viewers can follow the melody's rhythm while the emotional weight — already universal through music — transcends linguistic boundaries. As "Heaven Is a Place on Earth" swells over Kelly and Yorkie's digital eternity, the English lyrics feel not like a translation failure but an artistic choice.

Native Speaker Validation: The Human in the Loop

No translation system, however sophisticated, can validate itself. We submitted our completed subtitles to native Georgian speaker review, requesting detailed linguistic critique. The feedback validated our approach while catching subtle issues invisible to non-speakers.

A typo in line 38 — "უისმინე" instead of "მისმინე" (listen) — an extra character that spell-check cannot catch because both are valid Georgian strings. The name "Alison" transliterated as "ელისონი" when standard Georgian convention prefers "ალისონი." Most subtly, the phrase "No shit" translated as "რა თქმა უნდა" (of course) when the scene's sarcastic tone demanded "აშკარაა" (obviously) — a distinction requiring not just linguistic knowledge but cultural calibration to American casual speech patterns.

These corrections, applied programmatically, produced the final subtitle files. The pipeline had achieved its goal: Georgian translations that a native speaker could watch without once being pulled from the narrative by awkward phrasing or unnatural constructions.

"To have another language is to possess a second soul." —Charlemagne

The Complete Pipeline: From SRT to Screen

The final system processes subtitle files through a carefully orchestrated sequence: SRT parsing extracts entries with timing codes intact; scene detection segments content into dramatic units; strided windows provide translation context; Claude Opus generates Georgian text with explicit instructions for verb agreement and natural register; JSON parsing with regex fallback extracts results; lyrics filtering preserves musical moments; and progressive saving enables resume capability for the inevitable API interruptions.

The output integrates seamlessly with standard video tooling. A single mkvmerge command adds Georgian as a selectable subtitle track alongside the original English, preserving viewer choice while enabling cross-cultural viewing experiences. The technical vanishes; only the story remains.

Why This Matters: AI as Cultural Bridge

Georgian is spoken by roughly four million people — a linguistic island in the Caucasus, its speakers often underserved by global media localization efforts. Professional subtitle translation for a single episode costs hundreds of dollars and days of turnaround time. Our pipeline produces comparable quality in under an hour at a fraction of the cost.

But the implications extend beyond Georgian. The architectural patterns — scene-aware context windows, linguistically-informed prompt engineering, graceful fallback parsing, automated lyrics detection — generalize to any language pair where cultural and grammatical nuance matters. Basque, with its ergative-absolutive alignment. Finnish, with its fifteen grammatical cases. Navajo, with its aspect-rich verb system. Each presents unique challenges; each could benefit from pipelines designed with similar care.

In the end, watching Kelly and Yorkie find each other across decades and digital eternities, subtitled in flowing Georgian script, something remarkable happens: the technology disappears entirely. There is only the story, only the emotion, only two people — real or simulated, does it matter? — choosing each other against all odds. The subtitles have done their job: they have become invisible. And somewhere in the Georgian hills, a couple watches together, understanding every word.