Design a Multi-Lingual Subtitle Workflow

Problem statement

Design a subtitle and caption workflow that ingests, translates, synchronizes, versions, and delivers timed-text tracks for a streaming library containing thousands of videos, each requiring subtitles in dozens of languages. The platform must support professional and crowd-sourced translations, automatic speech-to-text generation, manual sync-offset corrections, dynamic font styling, user track selection at playback time, and low-latency delivery to players worldwide.

This is not a simple file-upload problem. A two-hour film at 24 fps has approximately 172,800 frames. Subtitles must align to specific timecodes with tolerances below 40 ms to avoid visible drift. Each language variant is an independent artifact with its own version history, review status, and quality score. When a video is re-edited or re-encoded, every subtitle track must be re-synced. When a translator corrects a single line, the platform must propagate that fix without regenerating the entire file. When a viewer switches languages mid-playback, the player must fetch the new track in under 200 ms to avoid a visible gap.

Why the problem is distinctive

A generic content management system stores files and serves them. A subtitle workflow must manage temporal alignment across multiple representations of the same content. A subtitle file is not a blob; it is a structured sequence of timed cues, each with a start time, end time, text payload, optional positioning, and optional styling. The platform must parse, validate, merge, diff, and re-time these cues at scale.

The problem requires timed subtitle tracks in SRT and WebVTT formats, multi-language and user-submitted translations, automatic or manual sync-offset corrections, dynamic font styling, user track selection, scalability for thousands of videos with multiple languages, low-latency loading, version control, and auto-translation with manual editing.

The four architectural planes

1. Ingestion plane: upload, parse, validate, and normalize subtitle files in multiple formats (SRT, WebVTT, TTML, IMSC1, EBU-TT-D, CEA-608/708).
2. Translation plane: machine translation, human review, community contributions, quality scoring, and approval workflows.
3. Synchronization plane: timecode alignment, frame-rate conversion, drift detection, offset correction, and re-sync after video edits.
4. Delivery plane: CDN distribution, manifest generation (HLS/DASH), player negotiation, caching, and low-latency track switching.

A strong interview answer keeps these planes separate. It allows the translation plane to operate asynchronously without blocking delivery of already-approved tracks, and it permits the synchronization plane to correct timing without invalidating translations.

Public operating baseline

Netflix supports subtitles and captions in over 30 languages across 190+ countries for its approximately 230 million subscribers. Netflix publishes Timed Text Style Guides per language and uses IMSC1 (Internet Media Subtitles and Captions) as its primary interchange format. YouTube processes over 500 hours of uploaded video per minute and provides auto-generated captions in over 100 languages using speech-recognition models. BBC iPlayer delivers subtitles conforming to EBU-TT-D and IMSC1 profiles under Ofcom accessibility requirements.

For capacity planning, this answer assumes a mid-size streaming platform with 50,000 videos in the catalog, an average of 30 subtitle languages per video, 200 million subtitle-track requests per day, and a 4x peak multiplier during new-release windows. Unless a number is tied to a citation, it is a stated design assumption.