AdBreak is an AI service that finds where a finished video can take an ad. It reads the file itself — scene changes, dark frames and fades, the ends of spoken sentences, silence, loudness and motion — and keeps only the moments where the story already pauses. Every break comes back with a time, a placement (pre-roll, mid-roll or post-roll), the kind of pause, a plain-language reason, a confidence score from 0 to 1 and the distance to the nearest speech, so an editor can accept or set aside each break before a marker goes out.
Breaks are delivered as SCTE-35 markers (aligned to a frame the player can cut cleanly), HLS markers or VMAP. The same reading of the speech produces SRT and WebVTT subtitles, translated into 22 languages. Opening and closing credits and lower-thirds are detected and marked separately, and each stretch of picture is labelled with the scene, objects and people on screen.
Break rules are set with presets per program length: minimum confidence, time skipped at the start and end, the shortest gap between breaks, and the minimum and maximum number of breaks. Inputs include MP4, MOV, ProRes and IMF packages. Large libraries do not need manual uploads: originals are pulled from S3, FTP or HTTP, or sent through the HTTP API, and results are uploaded back the same way. Files are not stored; they are deleted after processing.
AdBreak does not insert, play or sell ads and does not package HLS or DASH streams; the existing player, ad server or SSAI system keeps that job. Pricing is per hour of video, starting at $0.70 / €0.70 and decreasing with monthly volume.
Key features:
- Natural ad break detection from picture, speech and sound (scene changes, black frames/fades, sentence ends, silence, loudness, motion)
- Per-break time, placement (pre/mid/post-roll), kind of pause, reason, confidence 0–1, distance to speech
- SCTE-35 markers aligned to a cleanly cuttable frame, HLS markers, VMAP
- SRT and WebVTT subtitles from the same speech pass, translated into 22 languages
- Opening/closing credits and lower-thirds detection, marked apart from breaks
- Scene, object and people labels on each stretch of picture for quick human review
- Presets per program length: min confidence, start/end skip, min gap, min/max number of breaks
- Inputs: MP4, MOV, ProRes, IMF packages
- HTTP API; automatic pull of originals from S3/FTP/HTTP; results uploaded back
- Files deleted after processing (not stored)
- Human-in-the-loop: accept or set aside any break before markers go out
