On November 2, SceneMaker described work with SBS on 79 baseball clips reviewed by operators. The team says full evaluation is ongoing. These are developer-reported results that SIExpo has not reproduced. The project is associated with Solbox and SBSi in South Korea.
### From a long recording to an editorial proposal
The pipeline combines Qwen and pyannote for speech, text processing and retrieval. These are separate functions. A transcript describes what was said; timestamps help connect those words to a position in the recording. Proposing an excerpt is a further editorial step that still needs to be judged against the source footage.
The project’s public website describes four areas: finding highlights, creating shorter formats, matching relevant advertisements and processing archives in batches. Its displayed performance and quality targets are project objectives. They should not be read as results already achieved on every programme or as an evaluation of the new baseball workflow.
### Fewer frames provide a different view
In a separate preliminary test, the team reports faster processing at 0.5 frames per second but worse motion understanding. That is a material caveat for sports footage. Sparse sampling might preserve a clearly displayed caption while missing a brief action between selected frames that determines the meaning of an excerpt. This example is editorial explanation, not another measured result.
A pipeline can therefore be useful for finding candidate moments without being sufficient to decide their final significance. An editor needs to inspect the surrounding interval in the original recording. For baseball, that would mean checking what happens immediately before and after the apparent highlight as well as the action itself. Such a review helps avoid a clip that looks convincing alone but misrepresents the sequence of events.
### What is public, and what remains to be checked
The docs-web repository hosts SceneMakerAI’s documentation and technical publishing site. Its existence does not establish that the entire production application is public, or that every dependency shares a licence. A team considering a similar system should examine the models, code, video source and distribution rights separately. An open model does not automatically grant rights to broadcast footage.
For an editorial deployment, a sensible acceptance record would retain the proposed interval, the final edited interval and a reason for rejecting or changing a suggestion. Those are suggested evaluation practices, rather than features SIExpo has verified in SceneMaker. They would let a broadcaster distinguish a useful search assistant from a tool that merely generates many candidates.
Our editorial assessment is that the important connection is between each proposal and its original evidence. A useful follow-up evaluation would measure how many suggestions editors accept, how many they change and how long review takes. The number of generated clips alone cannot answer those questions. SIExpo has not measured these outcomes and does not present preliminary results as guaranteed savings for another broadcaster.
Editorial verification: 2026-11-02T09:49:31.133144+00:00 (UTC; Europe/Sofia is UTC+02:00).
