{"id":"b3810f9a-f51f-4de3-815a-7c41f9b8a816","arxiv_id":"2505.05654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A learned audio codec represents 2.98 seconds of audio as 32 sparse events, each decoded as a noise burst convolved with decaying resonances and a room impulse response.","lead":"This paper presents a proof-of-concept audio encoder that turns sound into a short list of sparse events, each with a timestamp and a learned vector that drives a physics-inspired decoder. A generalist might read it to see whether interpretable, event-based audio representations are a viable alternative to today's block-based and neural codecs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sparse-representation claim is unsupported by any fidelity measurement, and the paper's own §7 and §8.4 admissions undercut it; reconstruction quality must be quantified before the event codec can be evaluated.","rationale":"The reader's weakest assumption (decoder expressivity) identifies one possible cause of poor fidelity, but the more load-bearing gap is that fidelity is never measured. The paper's own admissions in §7 and §8.4 are in-scope evidence that reproductions fall short and events are redundant; these directly bear on the central claim. The fixed 32-step encoder also means sparsity is asserted, not demonstrated: it is a constant, not an outcome. I recommend no change to the conditional verdict because the reader already requires quantitative validation; however, I would state the condition more sharply: before any decoder-expressivity critique can even be evaluated, the paper must report objective reconstruction quality and an event-count/fidelity curve. The proposed test of reconstruction distortion versus event count would settle whether the central claim has empirical support.","tokens_in":6323,"tokens_out":5523,"duration_ms":57587,"concrete_test":"Run the released model on held-out MusicNet segments and report objective reconstruction quality after 1, 4, 8, 16, and 32 events: SI-SDR, PESQ, and log-spectral distance. Also count the number of events needed to reach a fixed distortion threshold (e.g., ≥15 dB SI-SDR or a chosen PESQ threshold). Compare against matching pursuit with the same number of atoms and the same dictionary/bit budget, and against a standard codec at matched bitrate. If the 32-event reconstruction is not clearly better than 32 matching-pursuit atoms, or if most segments require more than 32 events to reach the threshold, the central sparse-representation claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim—that audio can be represented as a sparse set of events—depends on 32 events per 2^16-sample segment being sufficient for perceptually faithful reconstruction. That condition is never tested. The paper reports no objective reconstruction metric (e.g., SI-SDR, PESQ, log-spectral distance), no listening test, and no comparison to matching pursuit or a standard codec. Its own text contradicts the claim: §7 states 'subjective reproduction quality falls short in this iteration,' and §8.4 says the model 'frequently produces redundant, duplicative events' while no sparsity penalty is imposed. In addition, because the encoder always runs a fixed 32 steps, 'sparse' is not an adaptive property: every segment, from a single tuning fork to a full orchestra, is encoded into exactly 32 events. Equation (1) therefore reports a fixed arithmetic ratio, not a measured compression property, and it ignores quantization of event vectors and times. The decoder-expressivity concern is real but secondary: until reconstruction fidelity is measured, no decoder design—physics-inspired or otherwise—can establish that the event representation actually represents the input.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a proof-of-concept audio codec in which an anti-causal convolutional encoder converts an STFT magnitude spectrogram into a fixed sequence of 32 events, each consisting of a coarse event time and a 32-dimensional event vector. A physics-inspired source-excitation decoder renders each event as noise bursts convolved with decaying sinusoidal resonances and a room impulse response, and the rendered event's spectrogram is subtracted from the residual. The authors argue that this yields a sparse, interpretable, and manipulable representation that could serve as an alternative to block-based codecs and as a substrate for future generative music models. The evaluation is entirely qualitative, via figures and online audio examples, and the text itself acknowledges that reproduction quality 'falls short' and that redundant events are common.","tokens_in":6547,"tokens_out":4035,"duration_ms":45449,"significance":"If the method lived up to its claims, a sparse event-based codec with physics-inspired rendering would be a genuinely interesting contribution to interpretable audio representation and could enable new forms of music manipulation and generation. The architecture is clearly described, the code is promised on GitHub, and the use of external room impulse responses as decoder priors is a sensible and well-motivated inductive bias. The paper is also honest about its limitations, which is commendable. However, the present manuscript does not establish the core claim: there are no objective reconstruction metrics, no listening test, no comparison to matching pursuit or standard codecs, and the authors' own conclusions concede both quality and sparsity problems. The strengths of the proposal and its presentation are not yet sufficient to support the title's claim of a working sparse and interpretable audio codec; the central claims are testable and could be supported by additional experiments.","major_comments":[{"comment":"The reported '~62x' compression is a ratio of sample count to scalar coordinate count, not a bit-rate ratio. It omits the bit-depth and quantization of both the 32-dimensional event vectors and the event-time scalars, and it is fixed by the architecture (32 steps for every 2^16-sample segment) rather than measured by any experiment. The claim of a 'compressive representation' is therefore not supported by this calculation. The paper should either report actual bitrates with a concrete quantization scheme or explicitly present Eq. (1) as an event-count sparsity measure rather than a compression ratio.","section":"Section 3.1, Eq. (1)"},{"comment":"The central claim that audio can be represented as a sparse set of events with good reproduction quality is never tested quantitatively. There are no reconstruction metrics (e.g., SI-SDR, PESQ, log-spectral distance), no controlled listening test, and no comparison to matching pursuit, MP3, Encodec, or even a simple spectrogram baseline, while Section 7 states that 'subjective reproduction quality falls short in this iteration.' Without such measurements, the paper cannot support the assertion that 32 events per segment are sufficient for perceptually faithful reconstruction, and the reader has no way to judge whether the representation actually represents the input audio.","section":"Sections 5 and 7"},{"comment":"The term 'sparse' is not an adaptive property of the representation: the encoder always runs a fixed 32 steps for every segment, and no sparsity penalty is imposed, so every input—from a single tuning fork to a full orchestra—is encoded into exactly 32 events. Section 8.4 concedes that the model 'frequently produces redundant, duplicative events that could be collapsed further.' This means the sparsity of the representation is an architectural constant, not an emergent or validated property. The paper should quantify how many of the 32 events account for most of the reconstructed energy, explore a threshold-based stopping criterion, and measure redundancy after collapsing near-identical events before claiming sparsity.","section":"Sections 4.2 and 8.4"},{"comment":"The source-excitation decoder, with its fixed final block and a single external room impulse response, is a strong modeling prior, and the paper offers only anecdotal evidence that this prior helps sparsity and quality. If this decoder is not expressive enough to render sustained instruments, vibrato, polyphonic textures, or unusual recording conditions, the representation will be incomplete for a large fraction of musical audio. At minimum, a controlled comparison with a standard convolutional decoder using the same encoder and loss would isolate the contribution of the physics-based design and give the reader a concrete test of the expressivity concern.","section":"Section 4.4"}],"minor_comments":[{"comment":"The manuscript contains numerous typographical errors, including 'intutive', 'paramaterized', 'enocder', 'manipulatibliity', 'snall', and 'analysed'; these should be corrected before publication.","section":"Throughout"},{"comment":"The text states that a window size of 2048 and hop size of 256 yield a 75% overlap, but the actual overlap is 87.5%; this should be corrected for accuracy.","section":"Section 4.1"},{"comment":"The symbol '∽' is used in place of the usual '≈' or '~' and is undefined; the authors should use a standard approximation symbol.","section":"Section 3.1, Eq. (1)"},{"comment":"The caption and surrounding text of Figure 3 appear to have formatting artifacts, including stray line breaks and mixed text ('N resonances', 'softmax', 'mixture dot product'); the figure and caption should be cleaned up.","section":"Figure 3"},{"comment":"Reference [5] is incomplete (missing author names and a full venue citation), and reference [7] contains inconsistent formatting for the conference name; the reference list should be completed and formatted consistently.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like an extended demonstration or position paper than a full archival evaluation: the architecture is interesting, but the absence of any quantitative reconstruction results and the authors' own admissions about quality and redundancy make the title's claim unsupported. I would support a major revision if the authors add objective metrics with baselines and a listening test; if the results then support the claims, the paper could be acceptable. The related-work coverage is fair and there is no sign of citation manipulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe take: this is a real architecture idea, not a demonstration that the idea works. The new thing is the combination—a learned anti-causal encoder that greedily picks one event per step, a source-excitation decoder with noise bursts, decaying resonances, and a fixed room IR, trained end-to-end on MusicNet. I haven't seen that exact package in the matching-pursuit/DDSP/spiking-autoencoder line it cites. The code and audio examples are promised, and the write-up is clear about what it does and doesn't do.\n\nWhat it does well: the decoder design is genuinely interpretable by construction, the streaming trick (encode first half, let events extend into the second half) is sensible, and the paper is honest. Section 7 concedes \"subjective reproduction quality falls short in this iteration,\" and Section 8.4 says the model \"frequently produces redundant, duplicative events\" with no sparsity penalty imposed. That candor is worth a lot.\n\nThe soft spot is the one the stress-test note names: the central claim—that sparse events represent audio—is never actually tested. There are no objective reconstruction metrics, no listening test, no baseline against matching pursuit or a standard codec. Equation (1)'s 62x figure is a scalar-count ratio, not a bit-rate ratio; it ignores quantization of event vectors and times, and because the encoder always runs 32 steps, \"sparse\" isn't even an adaptive property. A segment of silence and a segment of full orchestra both get exactly 32 events. Until reconstruction fidelity is measured, the paper supports \"here is a novel architecture\" but not \"this is a codec.\"\n\nThe decoder-expressivity worry is real but secondary. If the noise-burst-plus-resonances model can't render arbitrary instruments, the whole approach fails; but that's exactly what training should show, and the lack of metrics is the blocker.\n\nWho is this for? People working on interpretable audio representations, generative models over symbolic-ish audio, or differentiable DSP. It's a workshop-quality proof of concept with a clear path to a stronger paper. I'd send it to review—an editor shouldn't desk-reject a novel, honestly-written architecture paper—but the referee report should be firm about requiring quantitative evaluation and a corrected rate calculation before anything is accepted.\n\nMy verdict: interesting, not yet convincing. Worth a serious referee, not worth citing as evidence yet.","headline":"A genuinely novel proof-of-concept event codec with honest self-criticism, but no fidelity numbers yet; the sparse-representation claim needs measurement before it means anything.","tokens_in":7054,"tokens_out":1985,"would_cite":false,"duration_ms":20852,"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 new audio codec compresses music into a sparse set of timestamped events instead of fixed frames.","keywords":["sparse audio codec","event-based representation","source-excitation synthesis","interpretable audio encoding","matching pursuit","STFT magnitude spectrogram","neural audio compression","streaming audio encoder"],"falsifier":"Render an input that the source-excitation model should be unable to approximate, such as a sustained noise without decaying resonant modes, through the 32-step codec and inspect the residual spectrogram; if large irreducible energy remains after 32 events, the decoder's expressivity is the bottleneck and the proof-of-concept does not generalize to arbitrary audio.","tokens_in":6097,"feed_emoji":"🎵","tokens_out":4101,"duration_ms":43890,"temperature":0.7,"pith_summary":"This paper tries to establish that musical audio can be encoded as a short list of discrete events, each with a time of occurrence and a compact parameter vector, rather than as fixed-size blocks. If this works, a codec could give musicians an interpretable, editable representation: delete an event, shift it in time, or change the instrument and room that render it. The paper argues that a physics-inspired decoder, which renders each event as a noise burst shaped by decaying resonances and a room impulse response, encourages the encoder to find few events that still reconstruct the audio. It reports a proof-of-concept model trained on diverse classical recordings, with an approximate 62x compression ratio, while acknowledging that reproduction quality is not yet competitive.","feed_headline":"New audio codec stores music as sparse, editable events","feed_subtitle":"A physics-inspired decoder renders each event from a noise burst, decaying resonances, and a room impulse response.","key_machinery":"The load-bearing machinery is the source-excitation event decoder coupled with a greedy subtractive encoder. The encoder is an anti-causal dilated convolutional network that, at each of 32 steps, selects a single 32-dimensional event vector from the residual magnitude spectrogram and a coarse one-hot time position. The decoder block turns that vector into a noise burst, convolving it with a bank of exponentially decaying sinusoidal resonances, applying a time-varying mixture, and finally convolving with a fixed room impulse response; a scalar time-shift head provides sample-level scheduling. This physics-based inductive bias is what the paper credits for sparsity and for disentangling instrument resonance from room resonance.","core_discovery":"The central discovery is that an iterative encoder can decompose an STFT magnitude spectrogram into a small set of events, where each event is a scalar onset time plus a 32-dimensional vector, and a source-excitation decoder can render each event into audio by convolving a noise burst with decaying sinusoids and a fixed room impulse response. The encoder runs a fixed 32 steps; at each step it selects one event location, renders the event, subtracts its spectrogram from the residual, and repeats. The authors claim this representation is sparse and interpretable, exposes intermediate states such as impulse, resonance, and room, and supports operations like slicing, filtering, and shifting events that are awkward in block codecs. They present the result as proof-of-concept: subjective quality falls short, but the representation's properties warrant further study.","pith_inferences":["The paper does not impose an explicit sparsity penalty and notes that redundant duplicate events appear; adding an energy or sparsity term could reduce event counts further without architectural changes.","The event-vector space appears organized enough that nearest-neighbor search finds related variations, suggesting that semantic editing by interpolating event vectors could be feasible, an avenue the paper leaves unexplored.","Perceptual losses that tolerate small shifts in band-limited noise could allow the same fidelity with fewer events, since the current spectrogram loss spends capacity on perceptually irrelevant details.","If the source-excitation decoder generalizes beyond the trained classical corpus, the same representation could apply to any resonant or percussive sound, making event-based codecs a practical substrate for sound-design tools."],"forward_implications":["If the representation is correct, generative models could be trained to predict relationships between events instead of predicting frames, potentially improving long-term coherence in musical generation.","Editing audio becomes event-level: filter, delete, duplicate, or time-shift individual events rather than processing whole blocks.","Because the decoder separates impulse, resonance, and room response, one component can be swapped while holding others fixed, enabling separate manipulation of instrument and room.","A streaming encoder is possible because the model analyzes a longer segment while only scheduling events that begin in its first half.","Sparsity yields a rough compression ratio of about 62x at 32 events per 2^16 samples, making the representation potentially useful for storage and transmission.","The paper does not impose an explicit sparsity penalty and notes that redundant duplicate events appear; adding an energy or sparsity term could reduce event counts further without architectural changes.","The event-vector space appears organized enough that nearest-neighbor search finds related variations, suggesting that semantic editing by interpolating event vectors could be feasible, an avenue the paper leaves unexplored.","Perceptual losses that tolerate small shifts in band-limited noise could allow the same fidelity with fewer events, since the current spectrogram loss spends capacity on perceptually irrelevant details."],"supporting_citations":[{"why":"Supplies the sparse decomposition idea that this codec extends from many short atoms to fewer, longer events.","marker":"[5]"},{"why":"Provides the point-cloud-like event representation that the codec's events-in-time structure builds on.","marker":"[6]"},{"why":"Basis for the source-excitation decoder, modeling events as energy injected into a resonant system.","marker":"[11]"},{"why":"Provides the fixed room impulse responses used in the final decoder block, enabling disentanglement of room and instrument.","marker":"[12]"},{"why":"Supplies the training corpus of natural, diverse classical recordings used to demonstrate the proof-of-concept.","marker":"[13]"}],"fun_headline_variants":["Sparse audio codec stores sound as editable events","Physics-based codec renders audio from sparse events","Audio as sparse, interpretable events with physics model","Proof-of-concept codec: audio from sparse, editable events","Codec decomposes audio into sparse, physics-based events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on the assumption that a noise burst shaped by decaying resonances and a fixed room impulse response can render perceptually faithful events for any instrument and recording condition; if this source-excitation model cannot capture an arbitrary sound, the encoder's sparsity will come at the cost of unacceptable quality.","fun_headline_variants_meta":{"raw":{"variants":["Sparse audio codec stores sound as editable events","Physics-based codec renders audio from sparse events","Audio as sparse, interpretable events with physics model","Proof-of-concept codec: audio from sparse, editable events","Codec decomposes audio into sparse, physics-based events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1505,"prompt_tokens":847,"completion_tokens":658,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":580}},"tokens_in":463,"tokens_out":658,"duration_ms":6812,"temperature":1.0,"reasoning_tokens":580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:59:36.826073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render an input that the source-excitation model should be unable to approximate, such as a sustained noise without decaying resonant modes, through the 32-step codec and inspect the residual spectrogram; if large irreducible energy remains after 32 events, the decoder's expressivity is the bottleneck and the proof-of-concept does not generalize to arbitrary audio.","supporting_citations":[{"cited_title":"Matching pursuits with time-frequency dictionaries,","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse decomposition idea that this codec extends from many short atoms to fewer, longer events."},{"cited_title":"Xenakis, F ormalized Music: Thought and Mathematics in Composition","cited_arxiv_id":null,"evidence_quote":"Provides the point-cloud-like event representation that the codec's events-in-time structure builds on."},{"cited_title":"Fant, Acoustic Theory of Speech Production : With Calculations based on X-Ray Studies of Russian Articulations","cited_arxiv_id":null,"evidence_quote":"Basis for the source-excitation decoder, modeling events as energy injected into a resonant system."},{"cited_title":"V oxengo-ir,","cited_arxiv_id":null,"evidence_quote":"Provides the fixed room impulse responses used in the final decoder block, enabling disentanglement of room and instrument."},{"cited_title":"Learning features of music from scratch,","cited_arxiv_id":null,"evidence_quote":"Supplies the training corpus of natural, diverse classical recordings used to demonstrate the proof-of-concept."}],"review_version":1}