acceptodds
Under review as a conference paper at ICLR 2027

Transfer Asymmetry in Vehicle Pass-By Audio Synthesis: A Bidirectional Evaluation of PassByDiT and DopplerSim

Abstract

Synthetic audio offers a potential remedy for scarce labeled vehicle pass-bys, but acoustic realism alone does not establish its usefulness as training data. We study this distinction through transfer asymmetry in vehicle speed estimation. We introduce PassByDiT, a physics-conditioned diffusion transformer, and DopplerSim, a source-spectrum recovery and rendering pipeline, and compare both with Pyroadacoustics under matched pass-by conditions. PassByDiT achieves the closest agreement with real recordings across most spectral and temporal measures, while the physics simulators align energy peaks more precisely with the requested closest-point-of-approach timing. A real-trained speed estimator achieves similar mean absolute errors on real and PassByDiT audio ( and km/h, respectively), indicating that PassByDiT preserves motion cues usable by a model trained on real recordings. PassByDiT also provides the strongest synthetic-to-real speed-estimation performance among the evaluated synthesis methods. Bidirectional evaluation nevertheless reveals transfer asymmetry: agreement observed when testing a real-trained estimator on synthetic audio does not imply equivalent training utility in the reverse direction. Together, these findings establish complementary strengths of learned and physics-based synthesis and distinguish motion control, acoustic agreement, and training utility as separate dimensions of evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
vehicle pass-by audio, controllable audio generation, physics-conditioned diffusion, Doppler simulation, acoustic simulation, cross-domain evaluation, speed estimation, synthetic audio

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MDxbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the utility of synthetic vehicle pass-by audio for training downstream acoustic motion inference models (specific to vehicle speed estimators here), in response to the lack of motion-annotated real pass-by recordings. It introduces two complementary motion-conditioned synthesis frameworks, respectively PassByDiT, a physics-conditioned Diffusion Transformer that integrates 3-source kinematics and vehicle metadata into diffusion blocks, and DopplerSim, a physics-based pipeline that recovers an intrinsic vehicle source power spectral density (PSD) by inverting original Doppler and distance attenuation from a single reserved recording.

Critical strengths
  • The paper provides a clear and valuable conceptual distinction between acoustic realism, kinematic control, and downstream training utility.

  • The "transfer asymmetry" is insightful.

  • The methodology and evaluation look rigorous and solid.

  • The paper shows scientific honesty by emphasizing negative findings and limitations, as well as presenting detailed diagnostics.

Critical weaknesses
  • The experiment is entirely based on the VS13 dataset, which comprises only 13 vehicle models (all re-occur across splits) and 378 benchmark conditions. The generalization ability to unseen vehicles or conditions is uncertain.

  • Downstream utility is assessed only on scalar vehicle speed estimation using only the SE-ResNet architecture, which limits the effectiveness of the conclusions on other downstream tasks.

  • The benchmark protocol generates only one frozen random realization per condition with a fixed seed. Since diffusion models have inherent randomness, it would be better to explore the impact of sampling diversity.

  • While the diagnostic eliminates the frontend vocoder shift and flat ambient noise, why PassByDiT experiences a gap from 2.54 km/h in-domain MAE to 14.51 km/h on real audio remains essentially unknown.

Questions for authors

See weaknesses pls.

2 comments
AuthorsOpenReview ↗

Transfer Asymmetry Across Unseen Vehicles, Sampling Seeds and Training Conditions

Thank you for highlighting the transfer asymmetry. Shared concerns about unseen vehicles and the source of the gap are addressed in General Response G1 and G2 and summarized in points 1 and 4.

1. Generalization to unseen vehicle models

The original five-fold VS13 benchmark holds out recordings while vehicle models recur across splits. We ran a 13-fold leave-one-vehicle-model-out evaluation, excluding each held-out model from generator and estimator training and checkpoint selection (G1). Averaged over the 13 models, MAE was 8.11 km/h for real-to-real, 8.81 for real-to-PassByDiT and 15.95 for PassByDiT-to-real.

The paired directional gap was 7.13 km/h (95% vehicle-bootstrap CI [3.99, 10.71]), positive in 10 of 13 holdouts (two-sided exact sign test, p = 0.092). This supports average directional asymmetry across unseen vehicle models within VS13.

2. Scope of the downstream evaluation

Vehicle motion changes the recorded spectrum through the Doppler effect, with the shift depending on velocity relative to the microphone. Estimating speed tests whether generated clips preserve cues associated with the requested motion.

VS13 provides reference speeds and PassByDiT conditions on them. We use the same SE-ResNet architecture across training domains to compare audio sources under a consistent protocol. Our conclusions are limited to this task and architecture, which we state in the revision (G3).

3. Sampling diversity

We repeated the downstream evaluation with sampling seeds 1234, 2234 and 3234. Within each fold, we fixed the PassByDiT checkpoint, test conditions and requested vehicle motion. Each seed produced one clip per condition, keeping dataset size fixed. We retrained SE-ResNet on each resulting synthetic training set.

Sampling seedSynthetic to SyntheticSynthetic to RealReal to SyntheticReal to RealDirectional gap
12343.53112.5315.3625.6917.169
22342.16318.5617.1605.69111.400
32342.72417.6446.1435.69111.501

Columns list training audio first and test audio second. Each MAE pools 378 test predictions across five folds. All values are in km/h. The directional gap is synthetic-to-real MAE minus real-to-synthetic MAE. It remained positive for all three seeds, ranging from 7.2 to 11.5 km/h. Synthetic-to-real MAE varied from 12.5 to 18.6 km/h, so sampling affected the gap's size but not its direction in these runs.

We also varied the PassByDiT and SE-ResNet training seeds while fixing the sampling seed. Replacing real training audio with synthetic audio increased real-test MAE by 7.46 and 7.23 km/h in the two runs.

4. Why is there a gap between in-domain and real-recording MAE?

We cannot yet name a single cause. The diagnostics below narrow it down.

Frontend. In the original clip-holdout benchmark, PassByDiT-trained MAE was 2.54 km/h on synthetic audio and 14.51 on real recordings. On the same real test recordings, training on mel/Vocos-reconstructed real audio raised MAE from 7.05 to 8.05 km/h. Passing the real test recordings through mel/Vocos raised PassByDiT-trained MAE to 25.49 km/h (Appendix G). The reconstruction control increased MAE by about 1 km/h. The gap persisted when the reconstruction pipeline was shared.

Vehicle coverage (G2). We trained PassByDiT and its estimator on 2, 4, 7 and 10 vehicle models and tested on real recordings of held-out models. The real-trained curve was much flatter than the PassByDiT-trained curve (14.9 to 9.0 km/h against 47.7 to 17.7). The gap narrowed from 32.8 km/h with 2 models to 6.0 with 4, suggesting that limited vehicle coverage contributes strongly when few models are available. From 4 to 13 models the gap stayed around 5 to 9 km/h. It varied widely with which models were in training. We hypothesize that it would shrink further with more models than VS13 provides.

Direction. A real-trained estimator achieved 3.8 to 8.7 km/h MAE on PassByDiT audio across all vehicle counts and 3.8 to 4.1 once the generator had seen at least 4 models. Generated audio carries speed cues that a real-trained estimator can use. The reverse-direction loss is consistent with the PassByDiT-trained estimator relying on cues that are absent or different in real recordings.

Constant offset. The PassByDiT-trained estimator overestimates speed on real recordings. With 2 training models, mean error was +43 km/h against 47.7 km/h MAE. With 13 models (fold 0), mean error was +5.3 km/h and correlation with reference speed was 0.75, against 0.94 for real training. Removing the mean error reduced MAE from 13.0 to 12.2 km/h, about 0.8 km/h of the 7.6 km/h gap.

Per-vehicle variation. On the original real recordings, excess MAE relative to real training ranged from −0.34 to +20.46 km/h across vehicles and varied nonmonotonically across speed bands (Appendix I).

We will add these analyses, per-vehicle results, protocols and the vehicle-count figure to the revision.

Thank you for the detailed response and additional results. I appreciate the honest discussion on limitations. I would maintain my positive rating. Good luck!

Reviewer_KwFwOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper proposes PassByDiT for motion-conditioned vehicle pass-by audio synthesis and compares it with DopplerSim and Pyroadacoustics. The main finding is that acoustic similarity and precise motion control do not necessarily translate into effective synthetic-to-real transfer for vehicle speed estimation. However, the evidence supporting the paper's central claims is incomplete.

Critical strengths

The bidirectional transfer evaluation is useful and exposes an important distinction between acoustic similarity and downstream training utility. The comparison across learned and physics-based synthesis methods is also informative.

Critical weaknesses

1. The central acoustic claims lack direct perceptual validation.

Despite focusing on audio synthesis, the submission provides no accessible audio examples in the supplementary materials and reports no listening study. The claimed improvements in acoustic realism are therefore supported mainly by numerical metrics, which cannot establish whether the generated sounds are perceptually realistic or contain obvious artifacts. This is a substantial validation gap for the paper's main claims.

2. The synthetic training and test protocols are not adequately controlled.

The generator has seen the real recordings corresponding to synthetic training conditions, but not those corresponding to synthetic test conditions. This creates a potentially important distribution mismatch and may affect both in-domain performance and the reported transfer asymmetry. The paper does not isolate this effect. Combined with the limited evaluation on a single dataset containing only 13 vehicle types and the relatively incremental methodological contribution, the evidence is insufficient to support the broader claims.

Questions for authors

1.Can the authors provide accessible audio examples or listening-test results supporting the claimed acoustic improvements?

2.How do the main transfer results change when every synthetic training example is generated by a model that has not seen its corresponding real recording?

3 comments
AuthorsOpenReview ↗

Perceptual Validation, Cross-Fitting Control and Generalization Evidence

Thank you for the review. We conducted a listening pilot to evaluate perceived realism and a cross-fitting experiment to examine the effect of generator exposure to the corresponding real recordings. Generalization beyond the 13 vehicle models is a concern shared across reviewers, so those experiments are in the General Response (G1, G2) and summarized in point 3.

1. Perceptual validation

In a blind listening pilot, 10 participants each evaluated 12 clips from four audio sources across three matched vehicle conditions. The source of each clip was hidden. Listeners rated realism, reported audible artifacts, and guessed the speed category. We averaged realism ratings within each listener first.

Takeaway: listeners rated PassByDiT close to real recordings (4.37 against 4.47 out of 5) and above both physics simulators.

Audio sourceMean realism (1-5)Artifact reportedSpeed-category accuracy
Real4.4712/30 (40.0%)11/30 (36.7%)
PassByDiT4.379/30 (30.0%)12/30 (40.0%)
DopplerSim3.839/30 (30.0%)15/30 (50.0%)
Pyroadacoustics4.007/30 (23.3%)8/30 (26.7%)

PassByDiT received the highest realism rating among the synthetic methods (4.37/5, against 4.47/5 for real recordings), which is consistent with the acoustic metrics in the paper. DopplerSim led in speed-category accuracy (50%). This is a small pilot, and we will report it as such; the results motivate a larger blind study of how perceived realism relates to speed judgments.

We have also added matched audio examples from real recordings, PassByDiT, DopplerSim and Pyroadacoustics to the supplementary materials.

2. Cross-fitting control

When we generated each synthetic training clip with a model that had never seen its matching real recording, the speed estimator’s error on real test audio increased by 0.13 km/h in the evaluated fold. Transfer from synthetic to real audio still produced higher error than transfer in the reverse direction.

For each synthetic training clip, we excluded the corresponding real recording from the generator’s training and checkpoint selection. We then trained the speed estimator on the regenerated clips and compared it with an estimator trained under the original protocol, using the same test conditions.

Metric (km/h)Original protocolCross-fitted
MAE on real test audio after training on synthetic audio10.9111.04
MAE on synthetic test audio after training on synthetic audio3.823.24
Directional gap (synthetic-to-real MAE minus real-to-synthetic MAE)6.686.81

The +0.13 km/h difference was computed on the same 76 real test recordings (95% vehicle-bootstrap CI [−0.93, 1.16]). We will report the cross-fitting procedure and results in the revised manuscript.

3. Single dataset and 13 vehicle models

Unseen vehicles (General Response G1). The leave-one-vehicle-model-out evaluation is a stricter form of the control above: the generator and the estimator see no recording of the test vehicle at all. Across the 13 holdouts, mean MAE was 15.95 km/h when training on PassByDiT audio and testing on real recordings, versus 8.81 km/h in the reverse direction. The paired gap was 7.13 km/h (95% vehicle-bootstrap CI [3.99, 10.71]), with the same direction in 10 of 13 holdouts (exact sign test, p = 0.092).

Number of training vehicles (General Response G2). We also trained PassByDiT and its estimator on 2, 4, 7 and 10 vehicle models and tested on real recordings of the held-out models. The gap to real training on the same models narrowed from 32.8 km/h with 2 models to 6.0 km/h with 4 and stayed between 4.9 and 8.7 km/h through 10. We hypothesize that it would shrink further with more vehicle models than VS13 provides. An estimator trained on real audio had 3.8 to 8.7 km/h MAE on PassByDiT audio at every vehicle count, so the asymmetry does not depend on the number of training models.

Scope. All experiments use VS13, so our conclusions are limited to its 13 vehicle models and recording setup. We state this as a limitation in the revision (General Response G3). On the methodological contribution, please see point 1 of our response to Reviewer DAir.

Thanks for the additional experiments. The cross-fitting results suggest that generator exposure is not the main cause of the transfer gap, and the listening pilot provides useful initial evidence. However, two concerns remain:

  1. The methodological contribution is still limited. PassByDiT mainly applies a diffusion transformer with motion-related conditioning, and the rebuttal does not further clarify what is technically novel about this design.

  2. The generalization evidence remains limited. Despite the additional experiments, all results are based on VS13. The transfer gap also changes substantially with the number of training vehicle models, making it difficult to determine whether this finding generalizes beyond the current dataset.

Overall, I appreciate the additional efforts, but they do not change my Weak Reject recommendation.

We thank the reviewer for their response. We try to address both issues they have reiterated below, in reverse order. (Note that the second point is also the response we gave to reviewer 1).

Data:

The reviewer is right about the limited size of the data, but the fact is that VS13 is the only reliable dataset currently available. The reason for this is obvious -- simple safety considerations mean that it is dangerous to collect such data at scale.

One of the motivations for our model is in fact the shortage of data -- any modelling of vehicle dynamics requires data, which are not possible to collect, and must hence be synthesized, and we must do so leveraging only the very small real data available to us. To get around this limitation, we employ physics to ground the model, and the data to support it.

The true comparator here is other SOTA models for the same task, of which pyroadacoustics is currently the most popular. In all tests, including human subjective evaluation we beat it and other comparators by a significant margin.

Novelty:

Our methodological contribution is a pass-by-specific, frame-level motion representation that makes the temporal evolution of vehicle motion explicitly available to the diffusion model. While diffusion transformers, FiLM and attention-based conditioning are established techniques, conventional global conditioning does not explicitly encode how motion-dependent acoustic cues evolve throughout a pass-by.

Specifically, we derive eight frame-level signals from the requested speed and closest point-of-approach (CPA) time, representing front, center and rear CPA progression, amplitude evolution and Doppler variation. FiLM injects these signals into individual frames, while learned motion-cue affinity relates frames according to their motion-state similarity. This provides a structured connection between numerical motion specifications and temporally evolving acoustic generation, beyond vehicle metadata and global motion parameters alone.

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

This paper studies whether synthetic vehicle pass-by audio is useful for training speed estimators, focusing on transfer asymmetry between real and synthetic domains. The authors propose two synthesis methods: 1) PassByDiT, a diffusion transformer conditioned on motion features and vehicle metadata, and 2) DopplerSim, a physics-based pipeline that recovers source spectra from recordings and renders pass-by events under specified motion. Both are compared with Pyroadacoustics on the VS13 dataset, in terms of acoustic similarity, motion timing accuracy, and bidirectional cross-domain speed estimation.

PassByDiT matches real recordings best on most acoustic metrics and gives the strongest synthetic-to-real transfer. Models trained on synthetic audio perform substantially worse on real recordings than models trained on real data. The bidirectional results also show that strong real-to-synthetic performance does not imply equivalent synthetic-to-real utility, which means synthetic data should be judged on more than acoustic realism.

Critical strengths

The study systematically compares learned and physics-based synthesis methods across acoustic, temporal, and downstream task metrics. Five-fold cross-domain evaluation, 13-vehicle leave-one-out analysis, significance testing, and additional robustness checks make the evidence convincing. PassByDiT consistently matches real recordings more closely and gives the best synthetic-to-real speed estimation, while the physics-based simulators predict CPA timing more accurately. These complementary results shed useful light on the trade-offs between acoustic realism, physical controllability, and practical training utility.

Critical weaknesses

Although the synthesis framework is interesting, PassByDiT mainly combines established techniques such as diffusion transformers, FiLM conditioning, and physics-guided attention, and its methodological novelty beyond integrating them for this task is unclear. In addition, the conditioning ablations use a different dataset split and frontend from the main transfer experiments, so it is hard to tell how each architectural component contributes to synthetic-to-real generalization. Both issues weaken the case for the method’s significance.

Questions for authors

The main evaluation uses the same 13 vehicles in both training and testing partitions, and the leave-one-vehicle-out results are statistically inconclusive. Could the authors clarify how far the observed transfer asymmetry generalizes to unseen vehicles and different recording conditions?

1 comment
AuthorsOpenReview ↗

Methodological Contributions, Component Ablations and Generalization to Unseen Vehicles

Thank you for the constructive review and for recognizing the value of our bidirectional evaluation. We clarify the methodological contribution and the component-level evidence below. Generalization to unseen vehicles is a concern shared across reviewers, so the new experiments are in the General Response (G1, G2) and summarized in point 3.

1. What is new in PassByDiT (Response to Critical Weakness)?

Our methodological contribution is a pass-by-specific, frame-level motion representation that makes the temporal evolution of vehicle motion explicitly available to the diffusion model. While diffusion transformers, FiLM and attention-based conditioning are established techniques, conventional global conditioning does not explicitly encode how motion-dependent acoustic cues evolve throughout a pass-by.

Specifically, we derive eight frame-level signals from the requested speed and closest point-of-approach (CPA) time, representing front, center and rear CPA progression, amplitude evolution and Doppler variation. FiLM injects these signals into individual frames, while learned motion-cue affinity relates frames according to their motion-state similarity. This provides a structured connection between numerical motion specifications and temporally evolving acoustic generation, beyond vehicle metadata and global motion parameters alone.

2. What does each conditioning component contribute?

Appendix L (Tables 16 and 17) compares three variants on the same ablation split with the same audio frontend. Results are averaged over three training seeds. B conditions on clip-level speed, CPA time and vehicle specifications. F adds frame-level motion cues derived from the same speed and CPA inputs through FiLM. M adds learned similarity between frames to the attention logits.

Takeaway: the frame-level motion cues (F) are the component that matters. Learned affinity (M) adds little beyond them.

Acoustic metricBetterBFM
CPA-region L1Lower0.58480.47970.4800
Best-shift envelope correlationHigher0.79530.91370.9145
Log-mel L1Lower0.55840.51220.5117
MR-STFT distanceLower0.78890.75930.7585

Adding frame-level motion cues through FiLM produced the largest improvement on all four acoustic metrics. Adding learned affinity led to changes that were small relative to the variation across seeds.

The main five-fold benchmark uses a different split and audio frontend to evaluate the full model. PassByDiT achieves 14.51 km/h synthetic-to-real MAE, compared with 25.59 km/h for Pyroadacoustics and 35.73 km/h for DopplerSim. Measuring each component's contribution to downstream transfer would require evaluating B, F and M under the same transfer protocol. We will clarify the scope of both evaluations in the revised manuscript.

3. Does the asymmetry extend to unseen vehicles and recording conditions?

Unseen vehicles (General Response G1). We reran the leave-one-vehicle-model-out evaluation in both directions, excluding each held-out model from generator and estimator training and checkpoint selection. Mean MAE was 15.95 km/h when training on PassByDiT audio and testing on real recordings, versus 8.81 km/h in the reverse direction. The paired gap was 7.13 km/h (95% vehicle-bootstrap CI [3.99, 10.71]), with the same direction in 10 of 13 holdouts (exact sign test, p = 0.092).

Number of training vehicles (General Response G2). We also trained PassByDiT and its estimator on 2, 4, 7 and 10 vehicle models and tested on real recordings of the held-out models. PassByDiT-trained MAE fell from 47.7 to 17.7 km/h as models were added, while real-trained MAE on the same models fell from 14.9 to 9.0 km/h. The gap narrowed from 32.8 km/h with 2 models to 6.0 km/h with 4 and stayed between 4.9 and 8.7 km/h through 10. We hypothesize that it would shrink further with more vehicle models than VS13 provides. In the reverse direction, an estimator trained on real audio had 3.8 to 8.7 km/h MAE on PassByDiT audio, so the asymmetry is present at every vehicle count.

Recording conditions: VS13 uses one recording setup, so our transfer conclusions are limited to those conditions. We state this as a limitation in the revision (General Response G3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response on Transfer Asymmetry, Vehicle Coverage and Evaluation Scope

We thank the reviewers for their time! Two concerns were raised by more than one reviewer: whether the findings hold for unseen vehicle models and why a transfer gap exists. We ran two new experiments on these and report them here. The individual responses refer to this section.

In short: the transfer asymmetry holds for vehicle models never seen in training (G1). The gap depends strongly on vehicle coverage when training data from only a few vehicle models are available, falling from 33 to 6 km/h between 2 and 4 training models and then levels off at 5 to 9 km/h (G2).

G1. Unseen vehicle models: leave-one-vehicle-model-out evaluation

The original five-fold benchmark holds out recordings while vehicle models recur across splits. We reran the evaluation with 13 folds, each excluding one vehicle model from training and checkpoint selection for both the generator and the speed estimator. MAE is averaged over the 13 held-out models.

Takeaway: on vehicle models that neither the generator nor the estimator has seen, training on PassByDiT audio and testing on real recordings gives about 7 km/h higher error than the reverse direction. The asymmetry is not a result of vehicle models recurring across splits.

Estimator training audioHeld-out test audioMean MAE (km/h)
RealReal8.11
RealPassByDiT8.81
PassByDiTReal15.95

Synthetic-to-real transfer had 7.13 km/h higher MAE than the reverse direction (paired vehicle-bootstrap 95% CI [3.99, 10.71]). The gap was positive in 10 of 13 holdouts (two-sided exact sign test, p = 0.092).

G2. How does the gap change with the number of training vehicle models?

We trained PassByDiT and its speed estimator on 2, 4, 7 and 10 vehicle models and tested on every real recording of the held-out models. Each row averages six random vehicle subsets drawn from the fit and validation clips of fold 0. Every generator received the same number of optimizer steps. Every estimator was trained for the same 2,400 steps without early stopping, because early stopping halted fits on small subsets prematurely. All values are MAE in km/h.

Takeaway: real training is nearly insensitive to the number of vehicle models, while PassByDiT training is highly sensitive to it. The gap falls as the number of training vehicle models increases.

Training models (held out)PassByDiT-trained, on realReal-trained on the same models, on realGap (range over six subsets)PassByDiT-trained, on PassByDiTReal-trained (all 13), on PassByDiT
2 (11)47.714.9+32.8 (10.6 to 65.7)1.68.7
4 (9)19.113.1+6.0 (0.2 to 16.2)1.94.1
7 (6)15.610.7+4.9 (0.0 to 17.2)1.94.1
10 (3)17.79.0+8.7 (3.3 to 22.1)1.63.8

The first two columns are measured on real recordings of the held-out models. The last two are measured on PassByDiT audio for held-out conditions of the training models.

What this shows:

  • The real-trained curve is much flatter than the PassByDiT-trained curve. From 2 to 10 models, real-trained MAE falls from 14.9 to 9.0 km/h, while PassByDiT-trained MAE falls from 47.7 to 17.7 km/h.

  • The gap shrinks as vehicle models are added. It is 32.8 km/h with 2 training models and 6.0 km/h with 4. Limited vehicle coverage therefore accounts for most of the gap when few models are available.

  • From 4 to 10 models the gap stays between 4.9 and 8.7 km/h and the differences among these rows are within the variation across vehicle subsets. The leave-one-out result with 12 training models (G1: 15.95 against 8.11 km/h) and the 13-model benchmark (13.0 against 5.4 km/h on this fold) lie in the same range.

  • We hypothesize that the gap would shrink further with more vehicle models. At a fixed model count the gap depends strongly on which models are in training (0.2 to 16.2 km/h with 4 models, 3.3 to 22.1 km/h with 10), which indicates that it is sensitive to coverage. VS13 provides 13 models, so we cannot test this here.

  • The asymmetry is present at every vehicle count. A real-trained estimator reads PassByDiT audio at 3.8 to 8.7 km/h and at 3.8 to 4.1 km/h once the generator has seen 4 or more models.

The figure for this experiment will appear in the revised appendix.

G3. Scope

All experiments use VS13: 13 vehicle models, one recording setup, speed estimation as the downstream task and one estimator architecture. Our conclusions are limited to this setting. The revision states this in the limitations and avoids claims about other datasets, recording conditions, tasks or architectures.

Planned revisions

  • Leave-one-vehicle-model-out protocol and per-vehicle results (G1).
  • Vehicle-count experiment, its table and figure (G2).
  • Sampling-seed and training-seed experiments (response to MDxb).
  • Cross-fitting control and listening pilot (response to KwFw).
  • Clarified scope of the conditioning ablation versus the transfer benchmark (response to DAir).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.