acceptodds
Under review as a conference paper at ICLR 2027

Preserve, Then Bound: Conditionally Informative Acoustic Context for Few-Shot ATC Call-Sign Spotting

Abstract

Acoustic context in few-shot keyword spotting is neither uniformly irrelevant nor consistently reliable. Speaker and channel cues can disambiguate similar lexical targets while a local source association persists, yet become misleading when that association changes. We introduce , a framework that treats these cues as conditionally informative context while retaining lexical content as the primary evidence. summarizes class-stable lexical content with class prototypes and preserves transient acoustic context in an instance-level support memory. A pairwise relation network derives a candidate-wise contextual applicability cue for each query-candidate pair, using joint class- and source-agreement supervision during training without requiring source labels at inference. This applicability cue and margin-based lexical uncertainty jointly modulate a bounded contextual residual, allowing contextual evidence to refine uncertain lexical comparisons without granting it unrestricted decision authority. The bounded update preserves the lexical top-1 prediction whenever the lexical margin exceeds the maximum possible context-induced score difference. We evaluate on air traffic control call-sign spotting in ATCO2 and UWB-ATCC under class-disjoint 5-way 1-shot and 5-shot protocols with clean and acoustically degraded queries. Controlled ablations examine the contributions of instance memory, source-context supervision, contextual applicability, and lexical-uncertainty conditioning. By separating contextual applicability from decision authority, provides a strong lexically anchored recognizer and extends it with transient acoustic evidence, refining ambiguous comparisons while preserving sufficiently decisive lexical predictions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
few-shot learning, speech representation learning, metric learning, episodic learning, self-supervised representations, contextual representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_x1QtOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BioS2D, a framework for few-shot air traffic control (ATC) call-sign spotting that uses non-lexical acoustic information, such as speaker and channel characteristics, as auxiliary evidence. The framework uses a frozen WavLM-Large encoder with separate pathways for lexical and contextual information. Lexical embeddings are summarized as class prototypes, while contextual embeddings are retained in an instance-level memory. A learned applicability score and lexical uncertainty jointly control a bounded contextual adjustment to the final prediction. In 5-way 1-shot and 5-shot experiments on ATCO2 and UWB-ATCC, BioS2D outperforms the reported baseline systems, although most of the improvement comes from its lexical pathway. Compared with a matched lexical-only model, the contextual component provides small improvements on ATCO2, but no consistent improvement on UWB-ATCC or the harder 10-way diagnostic.

Critical strengths

One strength of this paper is its clear presentation of figures and tables, which help readers understand the proposed method and experimental results. The authors also report their results transparently, including cases where the proposed method provides little improvement or performs worse than the baseline. I appreciate this transparency, as it allows readers to better assess the actual effectiveness of the method.

Critical weaknesses
  1. The technical contribution appears rather limited. The proposed method uses a frozen WavLM encoder with two separate branches for lexical content and acoustic context. The lexical embeddings are averaged into class prototypes, while the contextual embeddings are stored individually and aggregated using LogMeanExp. The final prediction combines lexical similarity with a contextual score, controlled by a learned applicability gate and a hand-designed function of the lexical score margin. While this is a reasonable engineering approach, the individual components are relatively conventional. Furthermore, the scope of the problem is somewhat narrow, focusing specifically on few-shot ATC call-sign recognition. The method assumes that speaker and channel characteristics can provide useful clues about call-sign identity, but it is unclear whether this assumption holds for other speech recognition tasks. Overall, I do not see sufficient technical novelty or generality to justify an ICLR submission.

  2. The experimental improvements are marginal. As shown in Table 2, the proposed method improves accuracy by only 0.23 percentage points on ATCO2 (97.54% to 97.77%) and performs slightly worse on UWB-ATCC (98.52% to 98.49%) compared to the content-only baseline. Appendix Table 10 also shows that none of the improvements on the relation-balanced diagnostic remain statistically significant after multiple-comparison correction. Furthermore, the ablation results raise questions about the necessity of the proposed components. For example, replacing the instance-level context memory with a simple mean prototype achieves 97.75% on ATCO2, compared to 97.77% for the full model. Removing lexical uncertainty also produces nearly identical results (97.76%). On the more challenging 10-way diagnostic, the learned contextual correction even results in a slight performance decrease (−0.09 percentage points). These results suggest that most of the performance comes from the lexical branch, while the additional contextual components provide limited benefits.

Questions for authors
  1. Are the improvements over the content-only baseline in Table 2 statistically significant? Given the small differences and the ablation results, what evidence supports the necessity of the instance-level context memory and the lexical uncertainty mechanism?

  2. The applicability gate is trained using a target that requires both correct class identity and matching source context (Eq. 14). This means the gate may learn lexical class compatibility rather than actual source matching. Indeed, the class-controlled applicability AUROC is only 0.589, suggesting that the model has difficulty distinguishing matching and mismatching sources. Have the authors considered training the applicability network using source agreement alone? How much of the performance improvement can actually be attributed to the learned applicability gate?

1 comment

I appreciate the authors' detailed rebuttal and the additional experiments. The new results provide stronger evidence for the usefulness of contextual information, particularly the applicability gate and lexical uncertainty mechanism. However, I am still not fully convinced about the necessity of the instance-level context memory. Furthermore, the technical novelty of the proposed allocation mechanism remains limited, and its generality beyond this specific task is unclear. I therefore maintain my original rating, although I think this work could be a good fit for a speech-focused venue.

Reviewer_MdwWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents BioS2D, a framework for few-shot air traffic control call-sign spotting that treats non-lexical acoustics as conditionally informative context. It combines lexical class prototypes with an instance-level context memory and uses candidate-specific applicability estimates and lexical margins to control a bounded contextual residual. The resulting formulation provides a sufficient condition for preserving the lexical top-1 prediction. Experiments on ATCO2 and UWB-ATCC demonstrate strong overall recognition performance, while controlled diagnostics examine contextual information, its actual use, and its susceptibility to misleading source associations. The framework is coherent and the analyses are informative, although the contextual residual delivers limited realized gains.

Critical strengths
  1. Clear problem motivation and coherent method design. The paper addresses the conditional value of acoustic context: speaker and channel information can aid recognition while local associations hold, but may become misleading when those associations change. The principle of preserving information while limiting its decision authority has potential relevance beyond this application. Lexical prototypes and instance-level context memory reflect the distinction between stable class identity and transient associations. Candidate-specific applicability and lexical margins jointly control the contextual residual, yielding an explicit prediction-preservation condition.

  2. Extensive and transparent diagnostics. The relation-balanced tests, context-memory reassignment, class-controlled analysis, and oracle experiments help distinguish the information retained in context, the extent to which the model uses it, and its realized recognition benefit. These analyses also expose applicability estimation as a remaining bottleneck.

Critical weaknesses
  1. Limited realized benefit from the core mechanism and empirical support for some design choices. Most of the improvement over external baselines comes from the lexical pathway. Within the same trained model, enabling the contextual residual improves accuracy by only approximately 0.032 percentage points on the relation-balanced diagnostic and reduces it by 0.09 percentage points on the 10-way nearest-hard diagnostic. Instance memory also performs similarly to a mean context prototype, and source-relation discrimination reaches only 0.589 AUROC after controlling for class correctness. Moreover, none of the relation-balanced comparisons reported in Table 10 reaches the 0.05 significance threshold after Holm correction. Although the ablations indicate reduced sensitivity to reassigned context memories, the accuracy results provide limited support for the effectiveness of the contextual allocation mechanism or the necessity of all its components.

  2. A gap between the general framing and the evaluation scope. The experiments focus on short call-sign classification among enrolled candidates on two corpora. They provide limited evidence about source associations changing in actual temporal streams, and no cross-task evaluation establishes the broader applicability of the approach.

Questions for authors
  1. How can the inference-time contribution of context be isolated? Could the authors consistently report accuracy, correction rates, and harm rates with the contextual residual enabled and disabled within the same trained model across the main evaluation settings? How do these results explain the differences from the independently trained Content only control?

  2. What explains the gap between oracle and realized gains? How much of this gap is attributable to the applicability target combining class correctness with source agreement, the capacity of the estimator, or the quality of the contextual evidence? Would separately supervising source matching and class compatibility improve contextual allocation over the current joint target?

  3. How much does the full design improve over simpler fusion schemes? The existing Fixed applicability ablation already examines a simplified scheme that retains lexical-uncertainty gating. Building on this comparison, could the authors hold the lexical representation and training budget fixed and compare against tuned constant-weight residuals and margin-only gating? Do candidate-specific applicability and instance memory provide consistent gains across seeds and evaluation conditions?

  4. Does the method remain effective when source associations change over time? Could the authors evaluate temporally ordered episodes containing speaker or channel switches? Additional evidence from another few-shot task would also help establish the scope of the broader claim about conditionally informative context.

Reviewer_RutyOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes BioS2D, a framework for few-shot ATC call-sign recognition that combines lexical information with acoustic context. Lexical prototype matching provides the primary classification scores, while contextual evidence contributes a bounded residual adjustment. The adjustment is conditioned on estimated contextual applicability and lexical uncertainty, allowing context to influence ambiguous decisions while preserving predictions with sufficiently large lexical margins.

Critical strengths
  1. Clear and relevant motivation. Acoustic context can help distinguish lexically similar call signs, but its usefulness depends on whether associations between call signs and acoustic sources remain valid. The paper explicitly addresses this conditional reliability.
  2. Principled control of contextual influence. The bounded residual provides a sufficient condition under which contextual adjustments cannot change the lexical top-1 prediction. This offers a clear mechanism for limiting the influence of potentially misleading context.
  3. Informative experimental controls. The matched content-only baseline and controlled source-relation experiments help assess both the contribution of contextual modeling and its potential to introduce harmful shortcuts.
  4. Transparent reporting of variability and limitations. The proposed model and its ablations are evaluated across three independent training seeds. The paper also acknowledges the gap between contextual information identified by diagnostic analyses and the improvements achieved by the learned allocation mechanism.
Critical weaknesses
  1. The empirical benefit of contextual modeling is inconsistent. In Table 4, Content only achieves higher mean accuracy on ATCO2 1-shot and UWB-ATCC 5-shot. BioS2D improves accuracy in the remaining two settings by only 0.23 and 0.20 percentage points, respectively. These results provide limited evidence for a consistent advantage over the matched lexical baseline.
  2. The direct inference-time contribution of context is small. The 0.47-point improvement over Content only compares independently trained models and therefore cannot be attributed solely to contextual adjustments at inference. Within the trained BioS2D model, enabling the contextual residual increases relation-balanced accuracy from 97.998% to 98.030%, a gain of only 0.032 percentage points. The distinction between improvements arising from training and those produced by contextual allocation is therefore important.
  3. The harder diagnostic does not demonstrate improved recognition. On the 10-way diagnostic, BioS2D achieves 91.15% mean accuracy, compared with 92.01% for Content only. Separately, enabling the contextual residual within BioS2D reduces accuracy by 0.09 percentage points. Although the oracle analyses indicate potential improvement, the learned mechanism does not successfully realize that potential in this setting.
  4. Statistical evidence for contextual improvement remains limited. None of the seven comparisons displayed in Table 10 meets a 0.05 significance threshold after Holm correction over the stated family of 14 comparisons. For example, the 20% band-masking condition has an unadjusted p-value of 0.005 but an adjusted p-value of 0.073. The positive point estimates are encouraging, but the reported tests do not establish an improvement under the specified multiple-comparison correction.
  5. The evaluation supports a narrower application than operational call-sign spotting. The experiments primarily assess episodic classification of cropped recordings among enrolled candidates. They do not establish performance for continuous-stream detection, rejection of unenrolled call signs, or false-alarm control. Consequently, conclusions about practical ATC spotting should be limited to the evaluated setting.
Questions for authors
  1. Can the authors demonstrate a reliable benefit from the contextual residual itself, especially on the harder examples?
  2. How can applicability estimation be improved beyond lexical compatibility? Its class-controlled source-discrimination AUROC is only 0.589.
  3. Which UWB-ATCC classes have enough distinct recordings for five supports and five queries, and how are eligible classes selected?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.