Video unavailable. Open the source below.
NewMinusPod
A self-hosted podcast and YouTube ad remover whose native System One adapter asks one yes/no advertising question per transcript segment instead of asking a chat model where the ads are.
Video unavailable. Open the source below.
NewA self-hosted podcast and YouTube ad remover whose native System One adapter asks one yes/no advertising question per transcript segment instead of asking a chat model where the ads are.
One detection window becomes one request. The state is that window's transcript, one line per segment prefixed with a line id, and there is one noul question per segment keyed s<sid> that asks whether the line is advertising rather than editorial content. Returns are probabilities: a segment opens above the 0.95 detection threshold, extends at 0.40, and sits between them for review, and a category pass can add ordered choice questions over the same state.
Line L0001 of transcript is advertising, not editorial content.
The state is one detection window of the transcript, with each segment on its own line prefixed by its line id. A window produces one question per segment; two are shown, and overlapping windows are scored twice.
Line L0002 of transcript is advertising, not editorial content.
The state is one detection window of the transcript, with each segment on its own line prefixed by its line id. A window produces one question per segment; two are shown, and overlapping windows are scored twice.
A feed check finds an episode to process and its transcript is segmented, locally or through a configured transcription provider.
The segments become overlapping detection windows, and each window becomes the shared state of its own request.
Every segment gets its own yes or no advertising question, batched only by the optional questions-per-request and token budgets.
Confidence thresholds open, extend or drop an ad segment, and the boundary refinement and category passes can ask further questions before the cut is applied.
Ad detection becomes a set of independent calibrated probabilities with an explicit confidence gate, so the model never writes or explains and a borderline line is held rather than cut. Every batch for one window carries the same state, which keeps each question's context identical instead of letting batch boundaries change the answer.