🎬 Cue

A small multilingual model that scores subtitle break cost, plus a constraint-satisfying re-segmenter. SubER-first, guarantees normative compliance, and holds up on no-space languages (Japanese / Chinese / Thai).

SubER primary metric 0 hard-constraint violations borrowing time Netflix / BBC / custom profiles no ASR · no translation

💻 GitHub 🤗 Model 🤗 Dataset

What it does

Cue is a post stage: it takes an existing transcript with timing (SRT/VTT) and re-decides where lines wrap and cues break. A dynamic program chooses cue boundaries and in-line breaks together, minimising a learned break-cost while the hard constraints (line length, line count, reading speed / CPS, display duration) are removed from the search space — so every output satisfies the chosen profile by construction. It also does borrowing time: a cue claims adjacent silence to lower its reading speed.

Why it matters

Per-language break quality — Rule-based splitters collapse on no-space languages; Cue holds at 1.0.

CPS over time with and without borrowing — Borrowing time evens reading-speed density over the timeline.

Try it

The interactive Gradio demo needs a Gradio Space (HF PRO on free CPU). Run it locally in a few seconds:

pip install "cue-resegment[app,seg] @ git+https://github.com/NagaYu/cue"
python -m app          # opens the Gradio UI

# or from the command line:
cue resegment input.srt --profile netflix --lang ja

Or click ⋮ → Duplicate this Space and set the SDK to Gradio to host the live app.

Apache-2.0 · Cue is research/engineering software; the published model is a v0.1 proof-of-concept trained on a small sample (see the model card).