{"id":"7079805f-0fa9-42e1-a28e-3c0d1986f6e6","arxiv_id":"2509.07586","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A strictly causal, low-latency piano transcription system can reach 10-30 ms delays, but with a clear accuracy drop from removing lookahead and using shifted audio windows.","lead":"The paper adapts a state-of-the-art piano transcription model to run with minimum delay for live music, using only past audio instead of future audio. It finds that making the model causal and shifting the audio window cuts accuracy, but also exposes a hidden 10-second delay in the earlier system.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Comparing against an undertrained, reimplemented Mobile-AMT baseline may invalidate the tight-tolerance robustness claim; the SE-latency finding itself appears sound.","rationale":"The reader's weakest assumption—that 500/2000-epoch training is sufficient and that a reimplemented baseline is representative of the original Mobile-AMT—is the most load-bearing concern. The paper's headline finding about SE-layer latency is a structural argument that does not depend on training budgets, and the paper's own experiments show consistent trends across variants. However, the quantitative robustness advantage claimed at 10 ms is only as strong as the baseline it is compared against. An undertrained baseline could overstate Causal-AMT's relative performance. This concern is concrete and testable, but it does not overturn the paper's main architectural insight; it does make the empirical superiority claim conditional on a fair comparison. Hence the original CONDITIONAL verdict stands, and no change is needed.","tokens_in":11305,"tokens_out":3531,"duration_ms":40564,"concrete_test":"Retrain both models to convergence: run Mobile-AMT for the originally specified 3000 epochs on MAESTRO, and train Causal-AMT until its validation F1 plateaus (monitor learning curves; if 2000 epochs are insufficient, extend). Then recompute Table 4's onset F1 at 10/20/30 ms tolerances, using identical evaluation code. If Mobile-AMT's 10 ms F1 remains below Causal-AMT's within error bars, the robustness claim stands; if it rises above, the claim is an artifact of undertraining. Also report the onset threshold used for the final causal postprocessing and confirm code availability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—that Causal-AMT is more robust at 10 ms onset tolerance than Mobile-AMT—rests on Table 4, where Mobile-AMT is a reimplementation trained for only 2000 epochs, whereas the original Mobile-AMT is trained for 3000 epochs (Sec. 3.3). The paper justifies reduced training by saying 'most effects become evident already during the early training stages' (Sec. 4), but no learning curves or convergence checks are provided. If the original 3000-epoch model performs substantially better at tight tolerances, the F1 gap at 10 ms (31.55 vs. 18.26) could shrink or reverse. The final comparison also combines multiple modifications (binary targets, weighted loss, shifted asymmetric window, removed velocity conditioning, shared acoustic stack) without ablating the contribution of each to the robustness gain, and the final onset threshold is undisclosed. The paper's other central claim—that Mobile-AMT's SE global pooling over 10-second blocks introduces a hidden 10-second latency—is a straightforward architectural consequence and is not seriously threatened by training details.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper examines how the current state-of-the-art online piano transcription system, Mobile-AMT, can be adapted for minimum-latency real-time use. It identifies a hidden 10-second latency in Mobile-AMT caused by Squeeze-and-Excitation layers that globally pool over 10-second input blocks. The paper then proposes a strictly causal variant, Causal-AMT, combining causal convolutions, removal of SE layers, shifted asymmetric STFT windows, binary targets with weighted or shift-tolerant losses, causal postprocessing, and shared acoustic stacks. Experiments on MAESTRO v3.0 evaluate these modifications separately and in a final combined comparison, reporting that the causal model is more robust at tight onset tolerances (10 ms) despite lower overall F1.","tokens_in":11611,"tokens_out":3222,"duration_ms":37311,"significance":"If the SE-latency observation is correct, it is an important caveat for the real-time AMT community, as it exposes a flaw in a previously reported state-of-the-art latency figure. The paper provides a broad empirical exploration of preprocessing, label encoding, postprocessing, and architectural choices, and the authors release a baseline that could support future low-latency work. The paper also clearly identifies the fundamental difficulty of predicting low piano notes from less than one period of the fundamental at 10 ms latency. However, the central empirical claim of improved tight-tolerance robustness rests on a final comparison using a reimplemented Mobile-AMT trained for only 2000 epochs rather than the original 3000-epoch model, and no convergence checks are provided. This weakens the main quantitative conclusion, although the architectural latency finding is sound.","major_comments":[{"comment":"The central robustness claim (Causal-AMT F1=31.55 vs Mobile-AMT F1=18.26 at 10 ms) compares against a reimplementation of Mobile-AMT trained for 2000 epochs, whereas the original Mobile-AMT is trained for 3000 epochs (stated in §3.3). The justification in §4 that 'most effects become evident already during the early training stages' is not backed by learning curves or convergence checks. If the original 3000-epoch model improves at tight tolerances, the gap could shrink or reverse. Please train both models to the same budget (or use the original checkpoints) or explicitly reframe the comparison as equal-budget reimplementation results.","section":"§4.4, Table 4, §3.3"},{"comment":"The onset threshold used in the final comparison is undisclosed. Table 1 shows that onset threshold strongly affects F1, and that the benefits of different training losses depend on the threshold. Since the central claim is robustness at 10 ms, the threshold is a free parameter that must be reported, and ideally swept or chosen by a principled criterion applied identically to both systems. Without this, the comparison is incomplete and potentially unfair.","section":"§4.4, Table 4"},{"comment":"The paper states that 'across all experimental setups in Section 4.1 compared to the current one, all our causal modifications demonstrate significantly stronger robustness to decreasing tolerance thresholds.' However, Table 3 does not include the non-causal TP1 baseline from Table 1 under the same evaluation protocol, and the experimental setups differ in training segments, windowing, and postprocessing. This statement, and the similar claim in §5, should be supported by a direct comparison or qualified.","section":"§4.3, Table 3"}],"minor_comments":[{"comment":"The row labels 'TP1'...'TP5' are rendered ambiguously (e.g., 'TP19.58' reads as TP19). Use clear spacing or a separate column.","section":"Table 1"},{"comment":"The text says 'H2' for Hann 10 ms delay, but Table 2 labels it 'H2'; consider using H1/H2 consistently. Also, the claim of 'by about 20 dB' spectral leakage has no measurement details.","section":"§4.2"},{"comment":"The notation alternates between 'T5' and 'TP5'; please unify. Also clarify whether Tables 1–4 all use the same 3-second segment training except A6.","section":"§4.1"},{"comment":"Footnote 1 flags a missing detail in the reference Mobile-AMT about offset postprocessing. It would help to state whether this affects the comparison.","section":"§3.2"},{"comment":"The text 'BZ0' should be 'B♭0' (or 'Bb0').","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a music information retrieval venue. I do not see concerns about novelty or citation integrity. However, the main quantitative claim needs stronger empirical support: either longer training runs or a clearly stated equal-budget comparison. I would also encourage the authors to report the final onset threshold and consider a threshold sweep. The SE-latency contribution is solid and should be highlighted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main thing you need to know: the paper finds a real, previously unseen latency bug in Mobile-AMT. Its SE blocks global-pool over 10-second input chunks, so the actual online latency is 10 seconds, not the reported 174 ms. That's an architectural fact, cleanly explained. That alone is worth the read.\n\nThe rest is a systematic, honest study of how far Mobile-AMT can be pushed toward causal, low-latency operation. The individual techniques—binary targets, shifted asymmetric STFT windows, causal convolutions, shared acoustic stacks, causal postprocessing—are mostly known, but the paper evaluates them together at tight timing tolerances (10/20/30 ms), which is the right evaluation for the claimed application. They also release a baseline, though the code link is missing from this version, which is annoying.\n\nSoft spots are real. Most experiments use 500 epochs; the final comparison uses a reimplemented Mobile-AMT trained for 2000 epochs, not the original 3000-epoch model. The paper justifies shortening training by saying effects show early, but no learning curves or convergence checks are shown. So the headline robustness win at 10 ms (Causal-AMT F1 31.55 vs 18.26) is conditional on the baseline being undertrained. A fully trained Mobile-AMT could shrink or reverse that gap. In addition, the final model combines several modifications at once, so we don't know which one drives the 10-ms robustness. And the final onset threshold is undisclosed—a minor but fixable omission.\n\nThe SE-latency finding is the core contribution. It corrects a significant error in Mobile-AMT's reported latency and provides a causal baseline for the subfield. The rest is exploratory; the 10-ms goal remains out of reach, and the paper says so.\n\nFor peer review: yes, this deserves a serious referee. The architecture finding is important enough, and the experiments, though undertrained, are transparent. I'd ask for code, longer training or convergence curves, and disclosure of the final threshold. Not a desk reject. I'd bring it to a reading group if the group is into MIR or low-latency audio, and I'd cite the SE-latency point in my own work.","headline":"A useful system paper: it catches a real 10-second latency bug in Mobile-AMT's SE layers, but its headline 10-ms robustness win rests on an undertrained reimplemented baseline.","tokens_in":12040,"tokens_out":2610,"would_cite":true,"duration_ms":25676,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A state-of-the-art online piano transcription system, Mobile-AMT, carries a hidden 10-second latency from its Squeeze-Excitation layers, and this paper builds a strictly causal replacement, Causal-AMT, that fares better at 10 ms onset toler","keywords":["real-time piano transcription","latency","causal processing","Squeeze-Excitation","STFT window shift","onset detection","MAESTRO dataset","binary classification targets"],"falsifier":"Record the actual streaming behavior of Mobile-AMT: if it emits a usable onset prediction before receiving a full 10-second input block (for example by feeding a live stream and timing the first output), the claimed hidden 10-second latency is wrong; conversely, training Causal-AMT for the full 3000 epochs and showing it still trails Mobile-AMT at 10 ms tolerance would settle whether the robustness gain is real or an artifact of shorter training.","tokens_in":11234,"feed_emoji":"🎹","tokens_out":4858,"duration_ms":46771,"temperature":0.7,"pith_summary":"The paper aims to show that Mobile-AMT, currently the state of the art in online piano transcription, is not truly real-time: its Squeeze-Excitation layers globally pool over 10-second audio blocks, adding a hidden ten-second latency that the original latency calculation missed. The authors rework the system into Causal-AMT, a strictly causal model with shifted asymmetric STFT windows, binary onset targets, and causal postprocessing, and show that while overall F1 drops, the causal system is more precise and robust at tight 10–30 ms onset tolerances, which is what interactive musical use needs. The value is a credible open baseline for minimum-latency transcription plus a clear map of where the accuracy losses come from: preprocessing delay, causal convolutions, and label encoding.","feed_headline":"Causal redesign exposes hidden 10-second delay in piano AI","feed_subtitle":"Mobile-AMT's smoothing layers silently wait for a full 10 s of audio—the new causal baseline fixes that.","key_machinery":"The central mechanism is the combination of (1) replacing Squeeze-Excitation layers—whose global average pooling makes the model non-causal—with strictly causal convolutions whose receptive field extends only into the past; (2) a shifted, asymmetric STFT window that keeps a 2048-sample window for frequency resolution but ends only 160 samples (10 ms) after the reference point instead of centering on it, cutting the preprocessing delay from 1024 samples (64 ms); and (3) replacing triangular regression targets with binary pointwise targets plus weighted and shift-tolerant losses, so that detection and postprocessing need no future frames. The shared acoustic stack across onset/frame/velocity t","core_discovery":"At the paper's core is the claim that Mobile-AMT, the current state-of-the-art in real-time piano transcription, carries a hidden 10-second latency because its Squeeze-Excitation layers perform global average pooling over the full 10-second input block, a non-causal operation the original authors overlooked when reporting 174 ms. Starting from a reimplementation, the paper shows that making the model strictly causal—removing the SE layers, shifting the STFT window so it ends rather than centers on the prediction point (cutting preprocessing delay from 64 ms to 10 ms), using binary targets with a causal peak-picking postprocessor, and sharing acoustic-stack convolutions across targets—reduces","pith_inferences":["A likely next test is whether the accuracy gap closes with a longer training budget; the paper's 500-epoch screening and 2000-epoch final comparison leave open whether Causal-AMT's disadvantage is partly a training-duration artifact.","The hidden-latency audit generalizes beyond piano: any model using global pooling or non-causal normalization over a buffered input has latency proportional to buffer length, so the same check applies to other real-time audio ML systems.","A testable extension is an adaptive windowing scheme that grows the lookahead only for low pitches where the period exceeds the available samples, trading latency per note rather than globally.","Since the paper evaluates only algorithmic latency, measuring wall-clock throughput on edge hardware would determine whether Causal-AMT actually runs within a 10 ms frame budget end-to-end; the paper's own discussion leaves this to future work."],"forward_implications":["If Causal-AMT is taken as the baseline, researchers can now separate algorithmic latency (model causality) from engineering latency (inference speed, buffering) when designing real-time transcription systems.","The finding that shifting the STFT window to 30 ms delay matches centered-window performance while 10 ms hurts accuracy gives a concrete design curve for the preprocessing latency–accuracy tradeoff.","Removing velocity conditioning on onset prediction strongly improves onset accuracy, suggesting a simpler conditioning design for causal transcription models.","Sharing the acoustic stack across all targets shows near-parity, implying model size can be reduced without proportional loss, which matters for deployment on constrained hardware.","Evaluating at 10–30 ms tolerances rather than the standard ±50 ms better reflects interactive use, and Causal-AMT's higher precision at 10 ms indicates causal models can be better calibrated for real-time operation despite lower raw F1."],"supporting_citations":[{"why":"Mobile-AMT, the starting point and reference system; supplies the architecture being adapted and the reported 174 ms latency that the paper challenges.","marker":"[5]"},{"why":"Kong et al.'s offline high-resolution transcription model; supplies the base architecture and the triangular regression target encoding that the paper replaces.","marker":"[2]"},{"why":"MAESTRO dataset; the training and evaluation data on which all comparisons are run.","marker":"[1]"},{"why":"MobileNetV3; supplies the depthwise separable convolutions and Squeeze-Excitation blocks whose global pooling creates the non-causal latency.","marker":"[16]"},{"why":"mir_eval; supplies the note onset/offset F1 metrics used for evaluation across tolerances.","marker":"[17]"},{"why":"The shift-tolerant loss from beat tracking; used as a loss function for pointwise targets to handle timing tolerance.","marker":"[19]"},{"why":"Supplies the weighting scheme for positive onset/offset targets to counter label imbalance in binary classification.","marker":"[18]"}],"fun_headline_variants":["Hidden 10s piano AI lag fixed by causal redesign","Piano transcription: causal model cuts 10s delay to <30ms","Exposing the 10s non-causal wait in piano AI","From 10s lull to real-time: causal piano transcription"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The conclusions rest on the assumption that training each variant for only 500 epochs, and the final comparison at 2000 epochs against a reimplemented baseline originally trained for 3000 epochs, already reveals stable and representative differences; the paper gives no convergence check to confirm the rankings would survive longer training.","fun_headline_variants_meta":{"raw":{"variants":["Hidden 10s piano AI lag fixed by causal redesign","Piano transcription: causal model cuts 10s delay to <30ms","Exposing the 10s non-causal wait in piano AI","From 10s lull to real-time: causal piano transcription"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1070,"prompt_tokens":707,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":288}},"tokens_in":451,"tokens_out":363,"duration_ms":4254,"temperature":1.0,"reasoning_tokens":288,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:55:36.558658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the actual streaming behavior of Mobile-AMT: if it emits a usable onset prediction before receiving a full 10-second input block (for example by feeding a live stream and timing the first output), the claimed hidden 10-second latency is wrong; conversely, training Causal-AMT for the full 3000 epochs and showing it still trails Mobile-AMT at 10 ms tolerance would settle whether the robustness gain is real or an artifact of shorter training.","supporting_citations":[{"cited_title":"neither musical nor interactive","cited_arxiv_id":null,"evidence_quote":"Mobile-AMT, the starting point and reference system; supplies the architecture being adapted and the reported 174 ms latency that the paper challenges."},{"cited_title":"We will dis- cuss three noteworthy contributions [5–7]","cited_arxiv_id":null,"evidence_quote":"Kong et al.'s offline high-resolution transcription model; supplies the base architecture and the triangular regression target encoding that the paper replaces."},{"cited_title":"Exploring System Adaptations For Minimum Latency Real-Time Piano Transcription","cited_arxiv_id":"2509.07586","evidence_quote":"MAESTRO dataset; the training and evaluation data on which all comparisons are run."},{"cited_title":"Designing Neural Synthesizers for Low-Latency Interaction","cited_arxiv_id":"2503.11562","evidence_quote":"MobileNetV3; supplies the depthwise separable convolutions and Squeeze-Excitation blocks whose global pooling creates the non-causal latency."},{"cited_title":"On the relation between the fields of networked music performances, ubiqui- tous music, and internet of musical things,","cited_arxiv_id":null,"evidence_quote":"mir_eval; supplies the note onset/offset F1 metrics used for evaluation across tolerances."},{"cited_title":"Musical interaction at a dis- tance: Distributed immersive performance,","cited_arxiv_id":null,"evidence_quote":"The shift-tolerant loss from beat tracking; used as a loss function for pointwise targets to handle timing tolerance."},{"cited_title":"Im- proving networked music performance systems us- ing application-network collaboration,","cited_arxiv_id":null,"evidence_quote":"Supplies the weighting scheme for positive onset/offset targets to counter label imbalance in binary classification."}],"review_version":1}