{"id":"b11ed8ae-c8f4-4288-8c3c-f58fff8fa560","arxiv_id":"2607.19721","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 919-parameter CNN detects tropical bird vocalisations from log-mel spectrograms with 0.9810 AUC and 98.3% recall in under 7 kB of INT8 model size.","lead":"Researchers built a tiny bird-detector CNN that fits in under 7 kB and catches about 98% of bird calls in a Southeast Asian soundscape dataset. Because it runs on low-power microcontrollers, it could make long-term rainforest acoustic monitoring cheaper and more widespread.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 45-day storage claim compares Micro at α=0.10 to Goertzel's hit rate from α≈0.33; recomputing Goertzel at α=0.10 could erase or reverse the claimed storage gain.","rationale":"The reader's CONDITIONAL verdict is reasonable, and the central algorithmic results — 0.9810 AUC for a 919-parameter model, multi-seed stability, detailed ablations, INT8 degradation below 0.12%, and a public code/data release — appear well supported. However, the most load-bearing part of the abstract is the 'drop-in replacement' and '28→45 days' storage claim, and that is where a concrete internal inconsistency sits. Table 14 uses Benhamadi et al.'s Goertzel hit rate of 41.7% as if it applied at the same tropical prevalence α=0.10 used for DrongoNet-Micro, but with precision 0.71 and recall 0.90 that hit rate implies α≈0.329. This does not invalidate the architecture or the AUC/recall comparisons, but it means the headline storage-extension claim is not established as printed. A recomputation at a common prevalence, or a direct Goertzel measurement on SEABAD, is the single check that would settle it. I therefore keep the reader's CONDITIONAL verdict, but via a different concern than the label-noise audit highlighted by the reader.","tokens_in":29368,"tokens_out":8704,"duration_ms":88541,"concrete_test":"Recompute Table 14 at a common prevalence. Take Benhamadi et al.'s Goertzel trigger (or re-implement the AudioMoth Goertzel filter) and run it over SEABAD's positive and negative test clips to measure true-positive rate and false-positive rate at a recall-matched threshold. Then form p_Goertzel(α=0.10)=0.10·TPR+0.90·FPR and compare it to p_Micro=0.258. If p_Goertzel(0.10) ≥ 0.258, the claimed 28→45-day extension fails. As a minimal analytical check, verify whether the p=0.417 / PPV=0.71 / R=0.90 triple is consistent with α=0.10; it is not (it implies α≈0.329).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline deployment claim compares hit rates across different prevalences. Table 14 gives Goertzel recall=0.90, precision=0.71, hit rate=41.7%, and presents the 28-day Goertzel baseline as the comparison point for DrongoNet-Micro at α=0.10. But for any detector, precision = α·recall / hit_rate. With PPV=0.71, R=0.90, p=0.417, the implied prevalence is α = 0.71×0.417/0.90 ≈ 0.329, not the tropical 0.10 used for Micro. Holding α=0.10 fixed, the same p would imply Goertzel precision of only ~0.216, contradicting the quoted 0.71. Therefore the 41.7% / 28-day Goertzel row is not an α=0.10 operating point; the 'extends from ~28 to ~45 days' claim is computed by comparing Micro's hit rate at α=0.10 with Goertzel's hit rate at α≈0.33. The direction and size of the storage gain are unresolved: if Goertzel's FPR implied by its own operating point is carried to α=0.10, its hit rate can be near or even below Micro's, reversing the headline. Section 9's 'conservative for tropical soundscapes' argument does not fix this; it needs a recomputation at a common prevalence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DrongoNet, a family of three INT8-quantised CNNs for bird-presence detection on microcontroller-class hardware, trained and evaluated on SEABAD, a 50,000-clip Southeast Asian dataset. Through a four-phase ablation from a retrained TinyChirp baseline, the authors develop Nano/Micro/Edge (763/919/25,890 parameters; 5.09/6.26/33.06 kB; AUC 0.9727/0.9810/0.9991), with full-INT8 degradation below 0.12%. They report on-device latency and power on a Portenta H7 and project AudioMoth-class latency, SD-card storage life, and battery life, claiming that Micro is a drop-in Goertzel replacement at α=0.10 prevalence, capturing 8 pp more bird vocalisations and extending a 32 GB card from about 28 to about 45 days. The paper also reports transfer experiments to TinyChirp's Corn Bunting corpus, DCASE-2018, and BirdVox-DCASE-20k, and includes several negative results.","tokens_in":29727,"tokens_out":10629,"duration_ms":109513,"significance":"The measured core of the paper is strong: five-seed means with standard deviations, significance tests against the retrained TinyChirp baseline, a careful zero-shot/transfer analysis, and unusually candid reporting of negative results and limitations. The code and data release supports reproducibility. If the deployment claims survive a corrected common-prevalence comparison, the paper makes a useful contribution to TinyML bioacoustics by showing that sub-10-kB CNN gatekeepers can match a much larger baseline and potentially replace coarse energy triggers. The DCASE-2018 re-measurement, the 'in-region retraining' scope, and the explicit negative results on distillation and augmentation are valuable for practitioners.","major_comments":[{"comment":"The Goertzel reference row is not at α=0.10. From the tabulated Goertzel values, PPV = α·Recall / hit_rate, so α = 0.71×0.417/0.90 ≈ 0.33. Recomputing Goertzel at α=0.10 using the FPR implied by its own operating point (FPR ≈ 0.18) gives a hit rate of ≈0.25, versus Micro's 0.258; the claimed '~28 to ~45 day' storage gain essentially disappears and may reverse slightly. The 8 pp recall advantage is valid, but the card-life claim in the abstract and Section 10.2 compares Micro at α=0.10 with Goertzel at α≈0.33. Recompute all storage and battery comparisons at a common prevalence, and distinguish balanced-test precision (0.847) from field precision at α=0.10 (about 36% per Table 16).","section":"§7.3, Table 8"},{"comment":"As written, the operating threshold is selected by sweeping recall on the held-out test set (Table 8 reports test-set metrics), and the same test set is then used to report recall at the selected τ. The mean-recall floor is therefore a selection outcome rather than an unbiased performance estimate. If τ was actually chosen on validation, the text should say so and report that sweep; otherwise select τ on validation (or nested cross-validation) and re-report the test recall/FPR used in the deployment calculations. This is load-bearing because the recall–FPR pair at the operating point feeds directly into Tables 14–16.","section":"§10.4"},{"comment":"The paper itself states that residual false positives are dominated by insect stridulation, wind/vegetation rustle, and primate vocalisations, which are underrepresented in the negative training corpus. The deployment projections in Sections 9–10 use the balanced SEABAD test FPR=0.178; if field FPR is materially higher, the storage and battery numbers in Tables 14–15 degrade. The 'conservative for tropical soundscapes' argument only bounds Goertzel's FPR, not Micro's. Please add a sensitivity analysis of the claimed day/hour figures to plausible FPR inflation, or present the numbers explicitly as conditional on the SEABAD negative distribution.","section":"§9–§10"}],"minor_comments":[{"comment":"The Edge AUC is reported as 0.9986±0.0001 in Figure 5A but as 0.9991±0.0002 in Table 7, Table 10, and Section 8.2. Reconcile this inconsistency.","section":"Figure 5 vs Table 7"},{"comment":"The surname Benhamadi/Benhammadi is spelled inconsistently (e.g., Section 10.2, Eq. (2), and the reference list). Use one spelling consistently.","section":"Throughout"},{"comment":"The abstract says '8 pp more bird vocalisations'; at the stated operating points the recall difference is 0.983 − 0.90 = 8.3 pp. Use the precise value.","section":"Abstract and §10.2"},{"comment":"The caption claims Goertzel and DrongoNet-Micro 'operate at matched high recall'; their recalls are 0.90 and 0.983, so 'matched' is inaccurate. Rephrase once the comparison prevalence is fixed.","section":"Table 14 caption"},{"comment":"The phrase 'drop-in replacement for the Goertzel trigger' overstates the evidence, since no on-AudioMoth firmware port is reported and the AudioMoth latency/battery numbers are projections from Portenta H7 measurements and Benhamadi et al.'s coefficients. The limitations section already acknowledges this; consider softening the abstract wording.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The prevalence mismatch in the headline storage claim is the main obstacle and is fixable by recomputation. The threshold-selection ambiguity also needs clarification. The empirical model-building and benchmarking work is solid and should not be rejected for these defects."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the ML engineering is genuinely careful: multi-seed experiments, a four-phase ablation, an honest transfer study, and code/data release. Second, the headline “45 days vs 28 days” storage claim is not the clean comparison the abstract makes it sound. The core TinyML result is real: DrongoNet-Micro, with 919 parameters and a 6.26 kB INT8 footprint, reaches 0.9810 AUC and 98.3% mean recall on SEABAD, matching a retrained TinyChirp baseline within 0.1 pp at 28x fewer parameters. The ablations are useful and mostly well-executed — the finding that depthwise-separable convolutions hurt at 16 mel bins, and the frequency-emphasis layer replacing batch norm for INT8, are new and concrete. The negative results (zero-shot transfer collapse, augmentation not compensating for capacity) are reported honestly, which is rare. The on-device measurements on Portenta H7 are a plus. The soft spot is the deployment-economics comparison. Table 14 compares Micro at 10% prevalence with Goertzel's observed temperate-survey operating point, whose precision/recall/hit-rate imply a prevalence around 33%. At a common 10% prevalence, Goertzel's hit rate would be roughly comparable to Micro's, so the storage-day gain shrinks or disappears. The paper's “conservative for tropical soundscapes” argument doesn't fix that; it just asserts the FPR would be higher in the tropics. That needs a direct recomputation at the same prevalence. Also worth flagging: the operating threshold is selected on the test set (Section 7.3), which is a mild leakage, and SEABAD's labels were audited on only 1,000 clips at 97.8% accuracy, with negative examples drawn from a few corpora that may not cover all tropical confounders. These are addressable rather than fatal. The battery-life projections are explicitly from Benhamadi et al.'s coefficients and the paper acknowledges no on-AudioMoth validation yet. Who should read this: people working on TinyML bioacoustics, specifically MCU-scale bird detection, will get value from the architecture comparison and ablation. It deserves a serious referee, but the deployment claims need to be reworked before acceptance. I'd bring it to a reading group, and I'd cite it for the architecture and ablation results, not for the battery/storage numbers.","headline":"DrongoNet's core TinyML results are solid, but the headline storage gain over Goertzel is built on mismatched prevalence assumptions and needs recomputation.","tokens_in":30238,"tokens_out":3212,"would_cite":true,"duration_ms":32130,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DrongoNet-Micro, a 919-parameter INT8 CNN, can replace the Goertzel trigger in tropical bird recorders, raising recall by 8 percentage points and extending a 32 GB card from about 28 to about 45 days.","keywords":["bird audio detection","passive acoustic monitoring","tropical bird vocalisations","microcontroller deployment","TinyML","mel spectrogram","INT8 quantisation","embedded neural networks"],"falsifier":"Deploy DrongoNet-Micro at tau=0.35 on an AudioMoth-class recorder for several weeks in a Southeast Asian forest, and hand-label every triggered clip plus a random sample of untriggered ones. If measured recall is well below 98% at ~10% prevalence, or if the fraction of windows written is far from 25.8%, the central deployment claim fails. A lab check: run the model on an independently human-audited tropical test set and compare AUC/recall with the SEABAD results.","tokens_in":29231,"feed_emoji":"🐦","tokens_out":7142,"duration_ms":70882,"temperature":0.7,"pith_summary":"Passive acoustic monitoring in the tropics wastes storage and battery because most recorded audio contains no bird calls. This paper tries to show that a binary bird-presence detector small enough for a microcontroller—a 919-parameter, 6.26 kB INT8 CNN called DrongoNet-Micro—can replace the Goertzel energy trigger in commodity field recorders, catching 8 percentage points more bird vocalisations at 10% prevalence and extending a 32 GB card from about 28 to about 45 days. The same network matches a retrained 25.6 k-parameter mel-spectrogram baseline within 0.1 pp of AUC, so the advantage is efficiency, not raw accuracy. The authors also show the architecture retrains to other regions and species, but does not transfer zero-shot, which is why they frame deployment as in-region.","feed_headline":"6 kB neural net catches more tropical bird calls than field trigger","feed_subtitle":"Replacing Goertzel filters on AudioMoth-class recorders, a 919-parameter CNN keeps 98% recall in a 6.26 kB footprint.","key_machinery":"The load-bearing design is a four-phase ablation from a compact CNN baseline. For the microcontroller-size branch, the winning combination is global average pooling (removes the heavy classifier), focal loss (recovers the sensitivity lost to pooling by focusing on hard examples), and a 17-parameter learnable frequency-emphasis layer that scales each mel bin by sigma(w·s). That layer exists not for accuracy but because it gives batch-normalisation-like spectral normalisation without BN's instability under INT8 quantisation. Strided and depthwise-separable convolutions were tested and rejected at 16-mel resolution; standard 3x3 convolutions plus a 1x1 channel-mix stage survived. The operating","core_discovery":"The central claim is that at a six-kilobyte size scale, a binary CNN can do what a bird trigger needs: high recall on species-rich tropical soundscapes at a false-positive rate that saves storage rather than flooding it. On SEABAD—50,000 three-second clips, 1,677 species—DrongoNet-Micro achieves AUC 0.9810 ± 0.0016 and, at threshold 0.35, 98.3% recall at 84.7% precision. INT8 quantisation costs less than 0.12% AUC across the three variants. The smallest variant, Nano, sits at 5.09 kB and 0.9727 AUC; the largest, Edge, reaches 0.9991 AUC in 33.06 kB. Against a retrained TinyChirp CNN-Mel baseline, Micro is within 0.1 pp of AUC at 28x fewer parameters, and the difference is not statistically s","pith_inferences":["A field-deployable extension is to make the decision threshold adaptive to estimated local prevalence, since the paper shows precision collapses from 85% at balanced prevalence to 22-36% at 1:10-1:20; an operator could raise tau automatically under sparse conditions.","The paper's gatekeeper framing suggests a cascade: let DrongoNet-Micro run always-on as a wake-up trigger and gate a more expensive species classifier; this is implied by the design but not built here.","The learnable frequency-emphasis trick is a small, portable solution to INT8-safe spectral normalisation; it could plausibly improve other tiny audio models with narrow mel front ends.","The negative results on augmentation and distillation imply that for sub-1 kB detectors, collecting better regional training data will help more than algorithmic tweaks."],"forward_implications":["On an AudioMoth-class recorder, replacing Goertzel with DrongoNet-Micro at threshold 0.35 raises recall by about 8 percentage points and extends a 32 GB card from roughly 28 to 45 days at 10% bird-segment prevalence.","The full pipeline fits the device envelope: a 23 kB tensor arena (inside 32 kB SRAM), an inference time of 21 ms on a Cortex-M7, and a projected ~480 ms mel-plus-inference time on an AudioMoth-class Cortex-M4F—well inside the 3-second analysis window.","The architecture is deployable across acoustic domains once retrained on regional data: 0.9757 AUC on a Corn Bunting corpus at 28x fewer parameters than the dedicated baseline, and within 0.4 pp of a 14x larger model on DCASE-2018 in-domain.","Zero-shot transfer across domains does not work at this scale—Micro and Nano fall to near chance—so the paper's deployment claim is explicitly in-region and per-region retraining is required.","Full INT8 quantisation costs below 0.01% AUC for Nano/Micro and below 0.12% for Edge, so the tiny sizes are the real deployed sizes."],"fun_headline_variants":["919-parameter CNN at 6 kB out-detects Goertzel trigger on tropical birds","6 kB DrongoNet Micro nails 98% recall on 1,677 tropical bird species","Tiny 6 kB neural net replaces Goertzel trigger, catches more birds","DrongoNet Micro: 6 kB, 919 params, 98% recall for tropical bird AI","AudioMoth drop-in: 6 kB CNN detects more tropical birds than trigger"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The result rests on SEABAD's auto-curated labels (audited at 97.8% +/- 0.9% on 1,000 clips) being accurate and on the negative corpus covering the tropical sounds—insect stridulation, wind/vegetation, primate calls—that dominate the model's false positives. If label noise or negative-set gaps are systematic, the reported 0.9810 AUC and 98.3% recall are optimistic.","fun_headline_variants_meta":{"raw":{"variants":["919-parameter CNN at 6 kB out-detects Goertzel trigger on tropical birds","6 kB DrongoNet Micro nails 98% recall on 1,677 tropical bird species","Tiny 6 kB neural net replaces Goertzel trigger, catches more birds","DrongoNet Micro: 6 kB, 919 params, 98% recall for tropical bird AI","AudioMoth drop-in: 6 kB CNN detects more tropical birds than trigger"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001472,"raw_usage":{"total_tokens":5868,"prompt_tokens":970,"completion_tokens":4898,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":714,"completion_tokens_details":{"reasoning_tokens":4781}},"tokens_in":714,"tokens_out":4898,"duration_ms":34595,"temperature":1.0,"reasoning_tokens":4781,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:53:20.227879+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy DrongoNet-Micro at tau=0.35 on an AudioMoth-class recorder for several weeks in a Southeast Asian forest, and hand-label every triggered clip plus a random sample of untriggered ones. If measured recall is well below 98% at ~10% prevalence, or if the fraction of windows written is far from 25.8%, the central deployment claim fails. A lab check: run the model on an independently human-audited tropical test set and compare AUC/recall with the SEABAD results.","supporting_citations":[],"review_version":1}