lumi✦

Lumi / Features

YouTube Bilingual Subtitle Refinement & Smart Segmentation

Subtitle Refinement uses surrounding context to merge fragmented auto-captions into natural sentences before translating, while the Playback-first Refinement Window prioritizes the exact section you are watching right now.

Chrome ↗Edge ↗Configuration guide

Examples & specifications

Specifications & limits
ItemDetails
Refinement inputResegments and polishes existing platform caption text and timestamps (operates on text tracks rather than re-transcribing audio).
Deterministic translation chunksStandard translation slices ordered tracks deterministically from the start: up to 10 cues and 2,000 source characters per chunk.
Playback-first schedulingPlayback-first Refinement Window prioritizes your active playhead neighborhood and is distinct from Deterministic Subtitle Chunking.
Shared artifact matchingManaged artifacts are reused when input identity and Generation Profile Fingerprint match; text BYOK and video chat stay private.
Paired subtitle segmentation example
Before: 4 overlapping input cuesAfter: 3 resegmented cues
  1. 0.16–4.24 shi everyone so recently I gave a
  2. 2.28–6.40 s30-minute talk on large language models
  3. 4.24–8.48 sjust kind of like an intro talk um
  4. 6.40–10.08 sunfortunately that talk was not recorded
  1. 0.16–4.08 shi everyone so recently I gave a 30-minute talk on large language models
  2. 4.08–6.16 sjust kind of like an intro talk
  3. 6.16–8.20 sum unfortunately that talk was not recorded

Historical CLI sample recorded on 2026-09-27 from the same caption window (resegmentation only, translation disabled; the bilingual table below comes from a separate product demo), illustrating how fragmented cues merge into complete thoughts.

deepseek-v4-flash · temperature 0 · boundary-v1 · 0–180 s · en

Download sample and run settings
Recorded subtitle sample
TimeSourceLumi translation
21.60–26.00 sokay so let's begin first of all what is a large language model really好的，我们开始吧。首先，大语言模型到底是什么
26.00–29.76 swell a large language model is just two files right um嗯，大语言模型其实就是两个文件，对吧
29.76–32.60 sthere will be two files in this hypothetical directory在这个假设的目录里会有两个文件

Excerpt from the live product demo in its original English-to-Chinese language pair, illustrating how Lumi aligns and translates real cues. Andrej Karpathy · Intro to Large Language Models · 2026-06-27

Workflow

  1. Open a YouTube video and select your source caption track, target language, and AI model.
  2. Turn on Subtitle Refinement to immediately process the window around your current playhead.
  3. Jump to any chapter or timestamp — the scheduler automatically shifts priority to your new playback position.
  4. Read clean bilingual sentences alongside the video and customize the model or prompt style whenever needed.

Questions & answers

Why do YouTube auto-captions still feel fragmented when translated line by line?
YouTube's automatic captions are sliced by short acoustic pauses every two or three seconds rather than by complete grammatical clauses. Translating each fragment in isolation locks in those broken boundaries, scrambling word order and pronoun references across languages. Subtitle Refinement reads a sliding window of neighboring cues, rebuilds complete sentences with natural punctuation, and translates them in context. Because refinement operates on the platform's existing text track rather than re-running speech recognition, we recommend replaying key terms, names, or code snippets if the source captions contain a typo.
If I scrub to the middle of a long lecture, do I have to wait for the first half to finish refining?
No. The Playback-first Refinement Window dynamically tracks your playhead and schedules the captions around your current position first, so you can jump straight to a mid-lecture chapter without waiting for a monolithic full-video job. Because LLM inference and provider rate limits still take a moment, newly seeked passages display standard captions for a few seconds before seamlessly upgrading to refined bilingual subtitles. Window scheduling only governs subtitle processing order and never alters video playback speed or audio.
How does a Deterministic Subtitle Chunk differ from the Playback-first Refinement Window?
They solve two different engineering problems. Deterministic Subtitle Chunk is the canonical unit for standard subtitle translation: starting from cue zero of an ordered track, versioned rules slice captions into chunks of at most 10 cues and 2,000 source characters with strict 1-to-1 cue alignment, independent of where anyone starts watching. Playback-first Refinement Window is a real-time priority scheduler that decides which time range of Subtitle Refinement runs first; shared refinement artifacts additionally verify both input identity and the full generation profile.
If I switch AI models or use my own API key (BYOK), will Lumi still hit shared artifacts?
Shared AI Artifact reuse applies exclusively to managed workflows where both the source input identity and the Generation Profile Fingerprint (which hashes the model, target language, prompt/protocol version, and output parameters) match bit-for-bit. Changing your model or language always produces a fresh fingerprint so stale outputs are never served by mistake. BYOK text translation, refinement, and video Q&A are excluded from the shared pool. Hitting a shared artifact costs 0 Lumi Tokens, while new managed generations deduct actual input plus output Tokens.