Physiological signals provide a direct view of how music affects the body, yet remain underexplored as a retrieval modality. Prior work largely maps such signals onto low-dimensional valence–arousal representations, which are subjective and limited as ground truth. We frame physiology-driven music retrieval as a cross-modal contrastive learning task, investigating whether stimulus identity can be recovered directly from electrocardiographic traces and respiration signals of 110 participants listening to 30 classical piano performances. Both physiological and acoustic modalities are summarised using distributional statistics with participant-wise normalisation. Analysis using Earth Mover’s Distance shows true stimulus–response pairs align more strongly than random pairings (88.7% vs 50%, p < 0.001), indicating that the response contains a track specific signature within participants. A contrastive model trained with InfoNCE generalises this across participants to achieve Recall@1 = 0.662, Recall@2 = 0.909, and MRR = 0.814 in a 9-way retrieval task (p < 0.005). Reverse (music-to-physiology) retrieval collapses to chance, indicating shared stimulus-specific responses rather than individual responses. Despite the music features inducing a modest influence, explaining 14.1% of the variance in the physiological response, the effect is consistent across individuals and is sufficient to support cross-modal retrieval directly using physiological queries.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

Using only wearable ECG (heart) and respiration sensor signals, the contrastive learning model correctly identifies which of 9 piano tracks a listener heard well above random guessing, and crucially, this holds for participants the model was never trained on.

Q23. Main review and comments for the authors (to be completed before the discussion phase). Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.

This is an interesting paper that uses relatively low-cost autonomic sensors (heart rate and respiration) to train a contrastive learning model (on correct and mismatched pairs of sensor data and audio features) from 88 training-set subjects, each of whom listened to 9 piano tracks. They were then able to use this sensor/audio embedding model to predict/rank which of the 9 tracks each of the 22 test-set subjects was listening to when recording their autonomic response data. Recall @ 1 is 0.662, which is significantly better than a random guess of 1/9 = 0.111. The fact that autonomic responses can predict song identity across users is a very cool finding.

The ablation study in section 5.1 shows that the model relies mainly on autonomic variability features (e.g., respiration amplitude, LF, and HF power, which were most important) rather than simple rate measures like heart rate, which barely affected performance. This was a useful part of the paper because it shows which autonomic features are most useful for this relatively simple song-identification task.

One of my favorite ideas is that we could flip around the results for diagnostic purposes, since a known piece of music reliably elicits an expected autonomic response. That is, music could serve as a controlled, non-invasive probe of the autonomic nervous system, where deviations from the expected response might reveal useful information about the test subject.

The two main shortcomings of this paper are corpus-size and the lack of time series analysis. The corpus size (9 tracks but with different loudness and tempo modulations) is extremely small, making this less of a "retrieval" study and more of a music cognition/perception study. But it is hard to collect user data for a study like this, so the results are still useful and meaningful.

Also, the authors use exclusively non-temporal summary statistics from both the sensor and audio signal time series data. I was surprised by this. The representations could be extended with low-cost temporal descriptors (e.g., autocorrelation at fixed lags, coarse per-segment shape features, sample entropy, or cross-modal lag features) that remain fixed-length and easy to compute.

Overall, I found this paper to be very interesting, creative, and educational.

Q25. Meta-review and final comments for authors (to be completed after the discussion phase)

Thank you for your submission to ISMIR 2026. The reviewers found this paper to be a novel, well-executed, and thought-provoking contribution that opens an exciting direction at the intersection of music information retrieval and physiological sensing. We particularly appreciated the strong empirical results, the creative use of contrastive learning, and the demonstration that physiological responses can support music retrieval across previously unseen listeners. The work was viewed as technically sound, clearly presented, and likely to inspire future research in both Music-IR and music cognition. The reviews provided several constructive suggestions regarding additional methodological details, clarification of design choices, and opportunities for future extensions.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

Physiological signals can be used to identify songs across subjects using distributions.

Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.

Authors have use distributions to avoid requiring temporal alignment, but it could be more justified why temporal features are not used wether it is from authors's results or it comes from related work.

In abstract it is stated "Music features explain 14.1%of the variance in physiological responses", but I can not found information about that until 5.3, I can not find and understand well with the information that is provided how it is obtained, I would give a clearer definition of what it is and how it is obtained.

In section 4.3, I think that the information is a bit disorganized, there are different values and explanations that would need more context for a clear understanding. While the are clear ideas: as hufflings reduces performance to change or the difference between training and test is small, they lack a narrative to pass from one statement to the other.

I would use and choose one word between vagal and parasymphtetic.

As this paper is only with 9-track playlists, and I know that it can be difficult to obtain more data, maybe there could have been a comparision in performance, comparing 2, 3, 4, 5 tracks..., and to see how contrastive learning evolves with the presence of more classes.

Authors could provide the code that they have used to analyze Heart FM so they contribute more easily to reusable insights.

In the introduction I would also add if possible and space allows a mathematical definition of the metrics used RMSSD and SDNN to understand with more detail the meaning of the metrics.

In methods, if the data used was not recorded by the authors, if it was not recorded by the author's I think they delete that detailed technical and hardware explanations (as using the Polar H10....)

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

"Slow" physiological features can identify and retrieve corresponding music.

Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.

Strengths:

1. The paper is well-positioned for exploring physiological changes in listeners for music retrieval, especially as data collection at scale wouldn't require any additional behavioral overhead from the subjects, instead they could be tasked to just listen to the music.
2. The retrieval performance, especially with slow-varying physiological signals, is quite promising and merits further exploration of physiological MIR. Weaknesses:
3. The choice of using Wasserstein distances seems agnostic to the time-course of physiological changes during music listening. It would be interesting to see if circular permutations of the physiological features still successfully retrieve the music. Although the current baseline of shuffling labels makes sense, it limits retrieval to entire pieces rather than segments or sections within each piece.
4. Lines 181-182: Recruitment criterion for the subjects is not described. What is the age range, sex, hearing, and musical training of the subjects? I could not find this information in Ref. 35 either.
5. Lines 231-237: it is not clear what filtering was applied, especially as the physiological signals can be slow and contaminated with acquisition related offsets. Also it is not clear why both z-scoring and baseline normalization are used.
6. Figures 1 and 2 suggest that 8 compositions are used in this work, but the text notes 9. Also, since Fig.1 depicts realistic looking ECG, it might be better to show realistic breath inhalation/exhalation asymmetry rather than a sinusoidal breath.
7. The sectioning of the paper is confusing: Wasserstein distance results are reported in methods, ablation results are reported in discussion.

Other items that could be re-written or clarified:

• Line 134: RMSSD should be defined. Line 168 has a typo in sensors.
• Line 197: reproducing grand piano is confusing, perhaps the intent was a player piano?
• Line 202: recording rate and mono/stereo should be clarified.
• Line 203: "acoustic descriptors" and "music features" can be reconciled.
• Line 213: it would be helpful to clarify which clinical ECG lead this setup is closest to.
• Line 242: EMD can be confused with empirical mode decomposition.
• Line 253: time series duration (or if its the whole track used for z-scoring) can be clarified.
• Line 260: is "the participant heard" within-subject or inter-subject?
• Line 264: how many label shuffles contributed to the null distribution?
• Figure 2: what do the colors in the top (input) panel represent?
• Line 270: is chance 50% and not 1/8?
• Line 369: it might be helpful to motivate t-SNE here over PCA.
• Line 399: how embedding diagnostics were performed was unclear to me.
• Figure 4: it might be helpful to sort from high to low, and label the features with their expanded acronyms.
• Line 442: BH-corrected could be expanded.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

Similar physiological responses can be seen across subjects despite inter-subject variability.

Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.

Paper is strong, with a clear methodology, good understanding of physiology and computational methods, and good readability.

Based on the suggestions received from the reviewers and meta-reviewer, we have made the following changes to the manuscript:

1. Expanded the introduction and related work by incorporating additional references and discussion of prior studies.
2. Added participant recruitment criteria, demographic information, and further details of the recording equipment and acquisition protocol.
3. Included mathematical definitions of autonomic measures (e.g., RMSSD), clarified statistical measures, and expanded descriptions of the statistical analyses.
4. Improved the clarity of figures by expanding labels, updating the graphs and signal representation, and improving the formatting. The indentations were changed as per the ISMIR guidelines.
5. Reorganised and rewrote Sections 4.3 and 5.3 to improve logical flow, clarify the embedding diagnostics and variance analysis, and better motivate the reported results.
6. Revised the conclusion to better emphasise the broader implications of the proposed framework for physiology-driven music retrieval and future clinical applications.