Repetition as Free Supervision: Label-Efficient Ad Detection with Bidirectional Break-Level Context

Broadcast advertisements have an unusual property that most detection targets lack: they repeat. The same 30-second spot airs dozens of times across days and channels, while most program content airs once. Earlier systems use repetition as the detector itself, by matching incoming audio against past audio. That approach hits a hard ceiling, because roughly a quarter of ads air only once and can never be matched. We propose to use repetition in a different role: as a free, automatically mined training label. Segments that recur across weeks of unlabeled broadcast audio are overwhelmingly ads. A model trained only on audio to predict "this stretch recurs elsewhere in the archive" can then be asked to flag stretches that sound and sit like ads even when they never recur.

The second half of the idea explains why the model must look both backward and forward in time. A never-repeated ad is rarely acoustically distinctive by itself, but it almost always sits inside a commercial break, between other ads that do repeat. A model that reads several minutes of audio in both directions can see that an unfamiliar 25-second segment is flanked by known spots, and label it accordingly. A model that only reads the past cannot see the right-hand neighbor. So the note makes a two-part claim. Repetition-mined pseudo-labels can replace most of the 170 hours of hand labels used by the strongest audio-only system. Bidirectional context at the scale of a whole break is the ingredient that carries these labels over to the non-repeated tail.

Core Claim

The claim is that an audio-only detector pretrained on repetition-derived targets from unlabeled broadcast archives, then fine-tuned on about 5% of the labeled data, can match the error rate of a fully supervised pipeline. It should also detect a higher fraction of never-repeated ads. The gain on those ads should vanish when the model is restricted to past context or to a window shorter than a typical break. That last prediction is what separates this from "any pretraining helps."

Where This Sits

The closest audio-only system is Advertisement detection in French broadcast news using acoustic repetition and Gaussian mixture models. It chains a fast repetition search, detailed cepstral matching, Gaussian mixture classifiers for non-repeated ads, and a Viterbi decoding pass. Viterbi decoding is dynamic-programming alignment that picks the best global label sequence. The system reaches an ad detection error rate of 10.6%, with 96.7% recall and 93.0% precision, and it catches 87% of non-repeated ads. It motivates this proposal in two ways. It needs 170 hours of hand labels, and it reports that 25% of ads never repeat. We reuse its repetition search, but as a label generator rather than a detector.

The Viterbi pass already uses future context through a transition model between states. Our contrast is therefore not "bidirectional versus causal." It is a learned representation of several minutes of context versus a hand-set state machine over frame-level scores.

Two other systems constrain the design. Content-based advertisement detection shows that two-pass fingerprinting is fast, at 0.003× real time, but it only finds ads already in a catalog. BAF: An audio fingerprinting dataset for broadcast monitoring shows that fingerprinting collapses when music sits under speech, with F1 between 0.11 and 0.47. That result warns that our mined labels will be incomplete in exactly these conditions, and we measure this directly.

A Deep Neural Framework To Detect Individual Advertisement (Ad) From Videos finds candidate cuts at silences and compares ResNet-34 embeddings across each silence. It reports roughly 90% detection at 10% false positives, then relies on video features and a unidirectional LSTM to classify segments. We drop both the silence heuristic and the video stream.

For method, we borrow from two papers. Audio Mamba: Bidirectional State Space Model for Audio Representation Learning supplies the backbone. A state space model is a recurrent layer whose cost grows linearly with sequence length. Audio Mamba processes 80-second clips where the Audio Spectrogram Transformer runs out of memory at 20 seconds, but it was only trained with full supervision on short clips. Its authors name self-supervised pretraining for bidirectional state space models as open. Contrastive Bidirectional Transformer for Temporal Representation Learning supplies the strongest generic alternative objective. It predicts masked continuous features against distractors, which raised COIN segmentation accuracy from 34.3% to 53.9%. We treat it as a competitor, not as our method, because nothing in it targets the structure of commercial breaks.

Evidence from other domains motivates the bidirectional hypothesis. In sleep staging, transitional stages are resolved by what follows them (García Ciudad et al.), and the same holds for epidemic trend reversals (Song et al.). This is suggestive, not proof, for broadcast audio.

Research Program

Mining labels from repetition. We first collect several weeks of unlabeled, continuous broadcast audio from the same kind of Quebec French-language channels used by Gupta et al. We run their fast repetition search with the symmetric KL2 cepstral distance, plus detailed matching, across the whole archive. This yields a per-frame recurrence count and the number of distinct days and channels on which each segment recurs.

Raw recurrence is a noisy stand-in for "ad." Program jingles, news themes, and station idents also repeat. To separate them, we use features that come free from the archive. Ads recur at irregular times across channels, while program furniture recurs at fixed daily offsets on one channel. Segments that recur only at the same time of day get a separate "station element" target instead of the ad target. Recurrence under six seconds is known to be missed, so those frames are marked unknown rather than negative.

Pretraining the model. The model has two stages. A local encoder produces embeddings at about 20 ms resolution. A stack of bidirectional state space layers then runs over windows of 10 to 15 minutes, long enough to span a full break plus the program on both sides. The pretraining loss asks the model, from audio alone, to predict three things per frame: the recurrence label, the station-element label, and the positions where recurring segments begin and end.

The critical design choice is how to stop the model from simply memorizing fingerprints of the ads it has seen. We split the archive by time: models are evaluated only on ads whose first airing falls after the pretraining period. We also randomly mask entire recurring segments at the input, so the model must infer their labels from the surrounding break rather than from their own sound. A secondary masked contrastive term in the style of the Contrastive Bidirectional Transformer is added only as an ablation.

Fine-tuning. The pretrained model is fine-tuned on hand-labeled subsets of the Gupta et al. training set at 1%, 5%, 20%, and 100% of its 170 hours. It outputs ad/program labels and boundary positions, followed by the same kind of Viterbi smoothing as the baseline so that decoding is held constant.

Baselines

The naive reference point is repetition search alone, reported by Gupta et al. at an ad detection error rate of 33.2%. Any method that does not clearly beat it has learned nothing beyond what the label miner already knows.

The strongest existing audio-only method is the full pipeline of Gupta et al., with an ad detection error rate of 10.6%, 96.7% recall, 93.0% precision, and 87% of non-repeated ads detected. It is trained on 170 labeled hours. We re-implement it on the same splits so that we can also retrain it at each reduced label budget; this matters, since its published number exists only at the full budget. Beating it means one of two things. Either we match or improve on 10.6% at the 5% label budget, or we raise non-repeated detection above 87% at equal overall error. Boundary error should also improve on the reported 207 ms average miss and 198 ms average overshoot.

The Content-based advertisement detection fingerprinting system serves as a catalog-bound reference on repeated ads only. It reports detecting 18% more ads than a commercial baseline. It has no reported error rate on the Gupta test set, so we run it on that set with a catalog built from the pretraining archive.

A generic-representation baseline tests whether our targets matter at all. Here, frozen Wav2Vec2 features, which Uro et al. found strong for broadcast turn boundaries at 94–95% accuracy, feed the same bidirectional sequence head. They are fine-tuned under identical label budgets, window lengths, and hyperparameter search budgets.

A generic-pretraining baseline tests the same question from another direction. It pretrains the identical architecture on the identical unlabeled archive with the Contrastive Bidirectional Transformer's masked contrastive objective instead of repetition targets, using the same compute budget.

Three ablations isolate the key ingredient: a causal version of our model, versions with context windows of 30 seconds, 2 minutes, and 10 minutes, and a version with future context capped at a lookahead of 0 to 120 seconds. Finally, a bidirectional transformer over 1-second frame embeddings checks whether the state space backbone matters. At that coarse rate a 15-minute window is only 900 steps, well within attention's reach. We expect the backbone choice to matter mainly for boundary precision at fine resolution, not for detection, and we report it that way.

Why not just…?

Why not just fine-tune a frozen pretrained speech encoder with a sequence head on the small labeled set? This is the most serious threat. Generic self-supervised speech features already separate broadcast turn boundaries well (Uro et al.). They may also carry enough production-style cues, such as compression, music beds, and voice-over timbre, to flag ads with few labels. Our bet is that they carry no signal about break structure and never saw the recurrence statistics of an archive. The Wav2Vec2 baseline under identical budgets settles it. If it matches us on non-repeated ads at 5% labels, the repetition-mining contribution is not supported, and we would report the simpler recipe.

Why not just keep the repetition + GMM + Viterbi pipeline, which already uses future context? Its Viterbi pass does look ahead, and it already reaches 10.6% error. Its weaknesses are the 170-hour label requirement and a context model limited to state transitions over static frame scores. The pipeline retrained at 1–20% of labels shows whether its accuracy survives label scarcity. The context-length ablation shows whether minutes-long learned context adds anything beyond what Viterbi smoothing, applied equally to all systems, already provides.

Why not just do generic masked contrastive pretraining on the archive? It worked for video segmentation (Contrastive Bidirectional Transformer for Temporal Representation Learning), and it needs no mining step. However, the file's own open problem is that generic pretext tasks do not target boundary structure. The matched-compute comparison on the same archive decides this. If contrastive pretraining ties on non-repeated ads, repetition targets add nothing and the claim fails.

Why not just run a causal model so it can stream? Many monitoring and verification uses are offline or tolerate a buffer. Even so, latency is a real cost. The lookahead sweep reports exactly how much future audio is needed to recover the non-repeated-ad gain. That gives deployers a concrete latency-versus-accuracy curve rather than a blanket claim that bidirectional models are better.

Decisive Checks

The primary evaluation is the Gupta et al. French broadcast test set. We report ad detection error rate, recall, precision, detection rate on never-repeated ads, detection of ads under six seconds, and average boundary miss and overshoot in milliseconds, all as curves over the four label budgets.

Three results would confirm the proposal. First, the repetition-pretrained bidirectional model matches the 10.6% error rate using about 8.5 hours of labels. Second, its advantage over the Wav2Vec2 and contrastive baselines is concentrated on never-repeated ads. Third, that advantage disappears in the causal model and in windows shorter than a break.

A secondary check measures the quality of the mined labels themselves against the hand-labeled training set, reporting precision and recall of "recurs across channels" as an ad label, with background-music regions reported separately. The BAF results suggest recall will drop there. We also report results on the Canadian broadcast set of Content-based advertisement detection if access allows, to test transfer to English and French audio from different channels.

The proposal is falsified if either generic baseline matches it on never-repeated ads at low label budgets, or if limiting the model to past context costs nothing.