SubFlow
Going global with video requires three things — subtitles, translation, dubbing — and all three are usually outsourced: expensive, slow, inconsistent, with creators still fixing timecodes line by line. SubFlow turns the whole chain into one automated pipeline: upload a video, GPU transcription produces timecoded subtitles, an LLM performs context-aware translation, TTS generates multilingual dubbing, and the result is burned in and exported ready to publish. The entire pipeline deploys privately — footage never leaves your domain, built for the confidentiality demands of MCNs and brands.
The problem we solve
- Outsourced subtitles / translation / dubbing: hundreds of RMB per video, turnaround measured in days
- Line-by-line MT subtitles: inconsistent terms, broken style — going global becomes going away
- Manual timecode correction: an hour of work for a 10-minute video
- Sending footage to third parties means unreleased content can leak
Core capabilities
GPU transcription with timecodes
faster-whisper on RTX 5090: upload and get precisely timecoded SRT, multilingual recognition, long videos processed in parallel segments.
Context-aware translation
Not line-by-line MT: the LLM translates with full context, glossary and style settings — names, brand terms and jargon stay consistent across the whole video.
Multilingual TTS dubbing
IndexTTS2 local inference, multilingual dubbing with optional voice cloning, automatically aligned to subtitle timecodes.
Burn-in & multi-format export
SRT / ASS subtitle files, or ffmpeg burn-in straight to a finished MP4, with configurable style templates.
Batch queue processing
Queue an entire channel or season; the GPU scheduler handles the rest with real-time progress.
Private deployment
The whole pipeline runs inside your control domain — the answer to MCNs' hard requirement of keeping unreleased footage confidential.
Architecture
A four-layer pipeline that fits on a single GPU workstation or a private cluster.
Console layer
Next.js console
Upload, job orchestration, glossary management, progress dashboard, export & download.
Scheduling layer
FastAPI + worker queue
Job splitting, GPU queue scheduling, retries, batch parallelism.
Inference layer
faster-whisper · LLM · IndexTTS2
Transcription (GPU-accelerated), context-aware translation (local LLM + Claude API premium tier), multilingual dubbing.
Compositing layer
ffmpeg
Subtitle burn-in, audio track replacement / mixing, multi-format export.
Tech stack
- faster-whisper (GPU transcription, production-ready on RTX 5090)
- Local LLM + Claude API (two translation tiers: batch / premium)
- IndexTTS2 (local multilingual TTS, already used in UEIC's film pipeline)
- FastAPI + worker queue (job scheduling)
- ffmpeg (burn-in & compositing)
- Next.js (web console)
Pricing direction
Early access
Free
Invited creators and studios, full features, priority feedback loop.
Pay as you go
Per minute
Transcription / translation / dubbing billed by video minutes.
Studio subscription
Monthly
Bulk quota + queue priority + glossary / voice asset management.
Private license
Annual
The full pipeline deployed in your server room — content never leaves.
Current stage: PoC → Alpha. Transcription, TTS and burn-in are production-proven in UEIC's own film pipeline (M0 ✅); M1 = self-serve web loop; M2 = first paying external studio.