Watch the result
The ads remain visible; their audio is muted. Programme sound returns after each selected break. This is offline AI analysis with technical review and timing corrections. Playback uses the reviewed timestamps; it is not live AI running in the player.
Ready
For the clearest comparison, play five seconds before a transition. Switch “Original sound” / “Ads muted” at the same position. “Jump to next ad” starts the next break immediately and returns to the first after the last. The hosted preview streams a compressed copy at the original resolution.Detected and reviewed mute intervals
Includes full-screen network promotions within breaks. Keeps educational Weather Wisdom, local-forecast transitions, programme instructions, closing credits and sponsorship acknowledgements audible.
What this demonstrates
- Both source files were downloaded in full, totalling 52 minutes 37 seconds. Originals are retained.
- A multimodal model received both video and audio in 27 two-minute-or-shorter clips. The team inspected whole-recording contact sheets and one-second boundary sequences, then used short audiovisual checks for ambiguous sections.
- Programme demonstrations and sponsor logos are not enough to trigger a mute. For example, a Motorola logo over a forecast stays audible; the separate Motorola phone commercial is muted.
- The hosted player applies the reviewed mute windows to streaming video. Switching to Original sound immediately restores the source audio. Local master files retain the original video packets.
Honest test result: the raw model output needed correction. Four of 27 clips had invalid timestamp ranges; it also missed a Motorola ad and confused some programme information/credits with promotions. The final files use reviewed corrections. This is not an independent accuracy benchmark, a trained detector, or evidence of live-TV performance.
Verification passed: all 12 sampled mute intervals were silent; 14 sampled programme sections retained their audio level. The local masters’ video packet hashes match the originals, and both complete masters decoded successfully. Hosted previews are separately encoded for streaming compatibility. These checks verify processing, not detection accuracy.
Evidence and reproducibility
Final mute timestamps · Raw model results (including errors) · Source links and SHA-256 checksums · Audio/video verification · Test notes
Model: Gemini 3 Flash Preview for the second pass; initial Gemini 2.5 Flash pass rejected for timing problems. No source captions, online ad timestamp lists, or human annotations were supplied as ground truth. technical review is not an independent human-labelled benchmark. No TV/IR/CEC device was controlled.