Pith. sign in

REVIEW 3 major objections 4 minor 28 references

For binaural sound source localization, input feature choice often matters more than model complexity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 21:46 UTC pith:4LCMSOVQ

load-bearing objection A genuinely useful feature-ablation study for binaural SSL, with one credible ranking table and a headline 'features beat complexity' claim that the experiments don't actually support. the 3 major comments →

arxiv 2511.13487 v3 pith:4LCMSOVQ submitted 2025-11-17 eess.AS cs.LGcs.SD

Systematic Evaluation of Time-Frequency Features for Binaural Sound Source Localization

Davoud Shariat Panah , Alessandro Ragano , Dan Barry , Jan Skoglund , Andrew Hines This is my paper
classification eess.AS cs.LGcs.SD
keywords binaural sound source localizationtime-frequency featuresinteraural level differenceinteraural phase differencespectrogramconvolutional neural networkout-of-domain generalizationhead-related transfer functions
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that the choice of time-frequency input features affects binaural localization performance more than model size. Using a deliberately small CNN, it shows that on in-domain speech, a compact two-feature input—interaural level difference plus interaural phase difference—achieves a mean angular error of 4.5° at zero elevation. On out-of-domain sounds with mismatched head-related transfer functions, the same small CNN requires a richer three-feature set—left/right phase spectrograms plus both interaural cues—to reach 8.4°, outperforming much larger Transformer-based models. If correct, this tells practitioners to prioritize feature design before scaling architectures.

Core claim

The central finding is that feature composition, not parameter count, determines robustness in binaural localization. In matched conditions, ILD+IPD alone saturates performance; adding more features yields no gain. In mismatched, out-of-domain conditions, gains appear up to the combination of channel phase spectrograms with ILD and IPD, which reduces mean angular error from 22.7° (ILD alone) to 8.4°. A 0.1M-parameter CNN using this feature set beats three recent Transformer-based models with 0.6M–26M parameters on out-of-domain data.

What carries the argument

The objects under study are four time-frequency representations—magnitude spectrogram, phase spectrogram, interaural level difference (ILD), and interaural phase difference (IPD)—fed as multi-channel inputs to a three-layer CNN. ILD and IPD encode the head-shadow and phase-delay cues of the duplex theory; the claim is that combining explicit interaural cues with raw channel spectrograms lets the CNN learn content-robust spatial representations.

Load-bearing premise

The claim that features matter more than model complexity assumes the Transformer baselines are comparable: that their poorer out-of-domain accuracy comes from their input features (fewer or different cues) rather than from their classification output format, training procedures, or different data, since no controlled complexity sweep is run.

What would settle it

Train the same small CNN with the Phase L/R + ILD + IPD feature set at several widths (e.g., 0.1M, 1M, 10M parameters) on identical training data and show that out-of-domain MAE stays flat; or re-train a Transformer-based regression model on the same feature set and show that it still underperforms the CNN. Either result would directly test whether features, not architecture scale, drive the reported gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • In-domain binaural SSL systems can rely on just ILD+IPD and keep the model small.
  • Out-of-domain systems should include phase spectrograms alongside both interaural cues for best generalization.
  • Model complexity is not the primary bottleneck; the informativeness of the input representation is.
  • The optimal feature set provides a practical recipe for building speech-independent binaural localizers.
  • Continuous-angle regression avoids the coarse resolution floor that classification-based approaches impose.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The comparison may conflate feature choice with output formulation: two of the Transformer baselines are classifiers with 5° bins, so their resolution is inherently coarser; a controlled test feeding the same rich features to a regression Transformer would isolate the feature effect.
  • The feature ranking could be architecture-agnostic, suggesting a testable design principle: when domain mismatch is expected, start with Phase L/R + ILD + IPD before increasing parameter count.
  • The out-of-domain gain may stem partly from HRTF mismatch rather than content mismatch; separately ablating HRTF mismatch and content mismatch would sharpen the practical guidance.
  • The paper's own future-work direction—multi-source localization—could reveal whether these cues remain sufficient when sources overlap acoustically.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents a systematic study of four time-frequency features — magnitude spectrogram, phase spectrogram, ILD, and IPD — and their combinations as inputs to a small CNN for binaural azimuth localization. Models are trained on a speech corpus (TSP) rendered through SADIE HRTFs and evaluated in-domain (TSP–SSL) and out-of-domain (SynBAD–Var and SynBAD–Fix). The key empirical claims are that (i) on in-domain speech, the two-feature set ILD+IPD suffices, and (ii) on out-of-domain content, the richer set Phase L/R+ILD+IPD generalizes best, yielding an MAE of 8.4° on SynBAD–Var. The authors further compare their 0.1M-parameter CNN with three larger Transformer baselines and conclude that feature choice 'often has a greater impact on performance than model complexity' (Abstract, §3.2, §4). The paper also promises release of code and datasets upon acceptance.

Significance. If evaluated solely as a controlled within-architecture feature ablation, the paper offers a useful and reasonably complete map of feature-set performance for a fixed CNN under both matched and mismatched HRTF conditions. The planned public release of the Binamix-based dataset and code would provide a reusable benchmark. The paper also uses a circular angular loss and evaluates across multiple HRTF subjects, which are positive methodological choices. However, the headline claim that feature design outperforms model complexity is not supported by the present experiments, because it rests on an uncontrolled cross-model comparison. The test-set-based feature selection and the absence of error bars in Table 1 further weaken the internal ranking. These issues are fixable, so the paper could become acceptable after a substantial revision and re-analysis.

major comments (3)
  1. [§3.2 / Table 2] The central claim in the Abstract and in §3.2/§4 — that feature choice 'often has a greater impact on performance than model complexity' — is not directly tested. Table 2 compares a 0.1M CNN using Phase L/R+ILD+IPD against Transformer baselines that differ simultaneously in architecture, parameter count, input features, output head (continuous regression vs. 5° classification), and presumably training protocol. On the in-domain TSP–SSL set, FA ViT, which uses the same ILD+IPD features, beats the CNN by 2.7° vs. 4.5°, which is direct evidence that model complexity does not hurt when features are matched. To support the headline claim, the authors should run a controlled complexity sweep of the same CNN with fixed features (e.g., varying width/depth) or, alternatively, train a Transformer with the same features and output head. The statement in §2.3 that the simple design 'ensures our resu
  2. [§2.5 / Table 1] The best feature set is selected post hoc from the test sets (SynBAD–Var and TSP–SSL), and Table 1 reports single-run MAE values without error bars, confidence intervals, or significance tests. For example, the key out-of-domain advantage of Phase L/R+ILD+IPD (8.4°) over ILD+IPD (10.7°) could lie within run-to-run variability, and the 0.1° differences among several top sets on TSP–SSL are almost certainly noise. The authors should report mean±std over multiple random seeds (at least 3–5) and, ideally, perform feature selection on a held-out validation set before evaluating on the test sets. Without this, the ranking in Table 1 and the claim that 'adding more features leads to considerable performance gains' (§3.1) are not statistically grounded.
  3. [§2.1 / Abstract] The out-of-domain test sets SynBAD–Var and SynBAD–Fix are generated from Binamix [20], a library from the same group, and use SADIE HRTFs, sharing the rendering and interpolation pipeline with the training data. This does not make the evaluation circular — the test content and HRTF subjects differ — but it limits the external-validity claim of 'diverse content' and 'generalization to diverse content' in the Abstract. To make the generality claim stronger, the authors should evaluate on an independently produced binaural corpus or explicitly frame SynBAD as a content shift (and perhaps a different HRTF) rather than a full domain shift. As written, the single-dataset-pair comparison in Table 2 cannot support the broad statement that the CNN is 'the most robust model for diverse out-of-domain content'.
minor comments (4)
  1. [Throughout] There are several typos and inconsistencies: 'binural' should be 'binaural' in §2.3 and §4; 'Binamix' appears as both 'Binamix' and 'Binamix' in the references; 'Database V ersion' in reference [22] has a stray space; in Table 2, the CI notation is inconsistent (e.g., '±3.8' lacks the degree symbol in the FA ViT row).
  2. [§2.2 / Eq. (1)] The STFT definition in Eq. (1) omits the context that N is the FFT length and the window w is a finite-length analysis window; a brief notational clarification would help. Also, the feature extraction parameters (window 25 ms, hop 10 ms, FFT length N) should be stated explicitly for reproducibility.
  3. [§3.1 / Fig. 1] Figure 1 is informative but does not include confidence intervals or error bars, nor does it consistently mark the best-performing feature set across sources. Adding compact significance annotations would improve interpretability.
  4. [§2.3] The architecture description is clear, but the sentence 'This simple design allows us to ensure our results reflect the quality of the features, not the complexity of the model' is a conceptual claim, not a design guarantee. Consider rewording to 'we adopt a fixed, low-complexity architecture to isolate feature effects' — this would be accurate.

Circularity Check

0 steps flagged

No circular derivation; feature ranking is a held-out empirical evaluation, not a reduction to its own inputs.

full rationale

The paper's central result—that certain time–frequency feature combinations outperform others for binaural SSL—is obtained by training a fixed CNN on a held-out training set and measuring MAE on distinct test sets (TSP–SSL, SynBAD–Var, SynBAD–Fix). No equation in the paper defines a feature's effectiveness in terms of the final metric, and no fitted parameter is renamed as a prediction. The comparison against Transformer baselines in Table 2 may be subject to confounds (different architectures, output heads, training protocols), but that is an experimental validity concern, not circularity. The paper's own datasets and libraries (Binamix, SynBAD) are cited self-references, but they serve as data-generation infrastructure rather than as load-bearing mathematical support for the feature-ranking conclusion. The feature set selection is based on test-set inspection, which is a methodological weakness, but it does not make the reported MAE values equal to the inputs by construction. The derivation chain is therefore self-contained with respect to the claims evaluated.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No theoretical derivation is present. The study's findings rest on domain assumptions about STFT features, binaural cues, HRTF rendering, and the representativeness of the chosen train/test splits, plus hand-set preprocessing and model hyperparameters. The central claim does not introduce new entities or fitted physical constants.

free parameters (5)
  • STFT window/hop length = 25 ms window, 10 ms hop
    Hand-set in Sec. 2.2; fixes time-frequency resolution of all input features; no sweep reported, so feature ranking could depend on this choice.
  • FFT length (N)
    Not explicitly stated; only window/hop are given in Sec. 2.2. Affects frequency resolution and input dimensions; withheld detail complicates exact reproduction.
  • CNN capacity (conv channels 32/64/128, FC 128, dropout 0.3) = 0.1M parameters
    Hand-chosen 'lightweight' design in Sec. 2.3. The paper's complexity comparison never varies this capacity with fixed features, so the relative effect of capacity is untested.
  • Azimuth wrap to frontal plane = wrapping to -90°..90°
    Sec. 2.1 maps azimuths to frontal plane, discarding front/back distinction; this preprocessing choice affects error values and may favor certain cues.
  • Training hyperparameters (Adam lr=0.001, patience=20, max 1000 epochs) = lr 1e-3, patience 20
    Sec. 2.4: standard hand-set values; all models trained with same schedule, but no sensitivity analysis.
axioms (5)
  • domain assumption STFT of a binaural signal contains sufficient information for azimuth localization.
    The entire study feeds CNN with STFT-derived features (Sec. 2.2); if STFT loses needed phase/level information, rankings would change. Standard but unproven here.
  • domain assumption Duplex theory cue complementarity (ILD high-frequency, IPD low-frequency) holds for the CNN's learned use of features.
    Sec. 2.2 motivates ILD/IPD using duplex theory; the model is expected to exploit these cues in that frequency-dependent manner.
  • domain assumption SADIE II HRIR/BRIRs rendered via Binamix's Delaunay interpolation are a valid proxy for binaural recordings.
    Sec. 2.1 generates all data with Binamix [20]; if interpolation or rendering introduces artifacts, both training and test conditions share them.
  • domain assumption Train/test HRTF subject mismatch is representative of real-world listener/HRTF variability.
    Sec. 2.1 uses different SADIE subjects for train and test; cross-subject HRTF generalization is assumed to indicate real-world robustness.
  • domain assumption The fixed train content domain (TSP speech) and out-of-domain SynBAD test content are sufficient to measure generalization.
    Sec. 2.1: conclusion about 'diverse content' is based on one speech training set and one out-of-domain test set; sampling of content types is limited.

pith-pipeline@v1.3.0-alltime-deepseek · 7795 in / 13901 out tokens · 136464 ms · 2026-08-03T21:46:11.859390+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Systematic Evaluation of Time-Frequency Features for Binaural Sound Source Localization." pith.science (2026). https://pith.science/paper/4LCMSOVQ

@misc{pith2026251113487,
  author       = {Pith},
  title        = {Pith review of: Systematic Evaluation of Time-Frequency Features for Binaural Sound Source Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LCMSOVQ}},
  note         = {Machine review of arXiv:2511.13487}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This study presents a systematic evaluation of time-frequency feature design for binaural sound source localization (SSL), focusing on how feature selection influences model performance across diverse conditions. We investigate the performance of a convolutional neural network (CNN) model using various combinations of amplitude-based features (magnitude spectrogram, interaural level difference - ILD) and phase-based features (phase spectrogram, interaural phase difference - IPD). Evaluations on in-domain and out-of-domain data with mismatched head-related transfer functions (HRTFs) reveal that carefully chosen feature combinations often outperform increases in model complexity. While two-feature sets such as ILD + IPD are sufficient for in-domain SSL, generalization to diverse content requires richer inputs combining channel spectrograms with both ILD and IPD. Using the optimal feature sets, our low-complexity CNN model achieves competitive performance. Our findings underscore the importance of feature design in binaural SSL and provide practical guidance for both domain-specific and general-purpose localization.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 3 linked inside Pith

  1. [1]

    INTRODUCTION Binaural audio captures spatial hearing cues by recording or synthe- sizing sound with two channels to replicate the auditory information humans use for three-dimensional localization. Binaural SSL lever- ages interaural cues such as time and level differences, along with spectral shaping from the outer ear, to determine the spatial position ...

    2077

  2. [2]

    Below, we detail the datasets, feature representations, arXiv:2511.13487v2 [eess.AS] 18 Nov 2025 model architecture, model training, and evaluation

    EXPERIMENTAL SETUP To investigate the impact of feature choice on SSL performance, we train multiple models using various feature combinations and systematically compare their performance on in-domain and out-of- domain data. Below, we detail the datasets, feature representations, arXiv:2511.13487v2 [eess.AS] 18 Nov 2025 model architecture, model training...

    Pith/arXiv arXiv 2025

  3. [3]

    Impact of Feature Representations Table 1 summarizes the MAE of all feature sets on the TSP–SSL (in-domain) and SynBAD–Var (out-of-domain) test sets

    RESULTS AND DISCUSSION 3.1. Impact of Feature Representations Table 1 summarizes the MAE of all feature sets on the TSP–SSL (in-domain) and SynBAD–Var (out-of-domain) test sets. On the TSP–SSL, single features including ILD, IPD, Mag L/R, and Phase L/R show the highest localization errors. Combining these features reduces the error. For example, the two-f...

  4. [4]

    CONCLUSIONS This study presented a systematic evaluation of feature design for binural SSL, emphasizing the significant role of careful feature se- lection in achieving robust and accurate localization across varied conditions. Through controlled experiments, we demonstrated that the choice of input features often has a greater impact on perfor- mance tha...

  5. [5]

    Jens Blauert,Spatial hearing: the psychophysics of human sound localization, MIT press, 1997

    1997

  6. [6]

    Sound source localization using hearing aids with micro- phones placed behind-the-ear, in-the-canal, and in-the-pinna,

    Tim Van den Bogaert, Evelyne Carette, and Jan Wouters, “Sound source localization using hearing aids with micro- phones placed behind-the-ear, in-the-canal, and in-the-pinna,” International Journal of Audiology, vol. 50, no. 3, pp. 164– 176, 2011

    2011

  7. [7]

    Binaural sound source localization in real and virtual rooms,

    Monika Rycht ´arikov´a, Tim Van den Bogaert, Gerrit Vermeir, and Jan Wouters, “Binaural sound source localization in real and virtual rooms,”Journal of the Audio Engineering Society, vol. 57, no. 4, pp. 205–220, 2009

    2009

  8. [8]

    Robust binau- ral localization of a target sound source by combining spectral source models and deep neural networks,

    Ning Ma, Jose A Gonzalez, and Guy J Brown, “Robust binau- ral localization of a target sound source by combining spectral source models and deep neural networks,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 26, no. 11, pp. 2122–2131, 2018

    2018

  9. [9]

    Advanced binaural sound localization in 3-d for humanoid robots,

    Fakheredine Keyrouz, “Advanced binaural sound localization in 3-d for humanoid robots,”IEEE Transactions on Instrumen- tation and Measurement, vol. 63, no. 9, pp. 2098–2107, 2014

    2098

  10. [10]

    End-to-end binaural sound localisation from the raw waveform,

    Paolo Vecchiotti, Ning Ma, Stefano Squartini, and Guy J Brown, “End-to-end binaural sound localisation from the raw waveform,” inICASSP 2019-2019 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 451–455

    2019

  11. [11]

    Deep learning-based speech specific source localization by using binaural and monaural microphone arrays in hearing aids,

    Peyman Goli and Steven van de Par, “Deep learning-based speech specific source localization by using binaural and monaural microphone arrays in hearing aids,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1652–1666, 2023

    2023

  12. [12]

    On improved training of cnn for acoustic source lo- calisation,

    Elizabeth Vargas, James R Hopgood, Keith Brown, and Kartic Subr, “On improved training of cnn for acoustic source lo- calisation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 720–732, 2021

    2021

  13. [13]

    Exploit- ing cnns for improving acoustic source localization in noisy and reverberant conditions,

    Daniele Salvati, Carlo Drioli, and Gian Luca Foresti, “Exploit- ing cnns for improving acoustic source localization in noisy and reverberant conditions,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 2, no. 2, pp. 103– 116, 2018

    2018

  14. [14]

    Sound event localization and detection of overlapping sources using convolutional recurrent neural net- works,

    Sharath Adavanne, Archontis Politis, Joonas Nikunen, and Tuomas Virtanen, “Sound event localization and detection of overlapping sources using convolutional recurrent neural net- works,”IEEE Journal of Selected Topics in Signal Processing, vol. 13, no. 1, pp. 34–48, 2018

    2018

  15. [15]

    Improved feature extraction for crnn- based multiple sound source localization,

    Pierre-Amaury Grumiaux, Srdan Kitic, Laurent Girin, and Alexandre Guerin, “Improved feature extraction for crnn- based multiple sound source localization,” in2021 29th Euro- pean Signal Processing Conference (EUSIPCO). IEEE, 2021, pp. 231–235

    2021

  16. [16]

    A sequence matching network for polyphonic sound event lo- calization and detection,

    Thi Ngoc Tho Nguyen, Douglas L Jones, and Woon-Seng Gan, “A sequence matching network for polyphonic sound event lo- calization and detection,” inICASSP 2020-2020 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 71–75

    2020

  17. [17]

    Binaural sound localization in noisy environments using frequency-based audio vision transformer (favit),

    Waradon Phokhinanan, Nicolas Obin, and Sylvain Argen- tieri, “Binaural sound localization in noisy environments using frequency-based audio vision transformer (favit),” inINTER- SPEECH. ISCA, 2023, pp. 3704–3708

    2023

  18. [18]

    Auditory cortex-inspired spectral attention modula- tion for binaural sound localization in hrtf mismatch,

    Waradon Phokhinanan, Nicolas Obin, and Sylvain Argen- tieri, “Auditory cortex-inspired spectral attention modula- tion for binaural sound localization in hrtf mismatch,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 8656–8660

    2024

  19. [19]

    Bast-mamba: Binaural audio spectrogram mamba transformer for binaural sound localization,

    Sheng Kuang, Jie Shi, Kiki van der Heijden, and Siamak Mehrkanoon, “Bast-mamba: Binaural audio spectrogram mamba transformer for binaural sound localization,”Neuro- computing, p. 130804, 2025

    2025

  20. [20]

    Mambavision: A hybrid mamba-transformer vision backbone,

    Ali Hatamizadeh and Jan Kautz, “Mambavision: A hybrid mamba-transformer vision backbone,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 25261–25270

    2025

  21. [21]

    Learning robust spatial rep- resentations from binaural audio through feature distillation,

    Holger Severin Bovbjerg, Jan Østergaard, Jesper Jensen, Shinji Watanabe, and Zheng-Hua Tan, “Learning robust spatial rep- resentations from binaural audio through feature distillation,” arXiv preprint arXiv:2508.20914, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    A survey of sound source localization with deep learning methods,

    Pierre-Amaury Grumiaux, Srdan Kitic, Laurent Girin, and Alexandre Guerin, “A survey of sound source localization with deep learning methods,”The Journal of the Acoustical Society of America, vol. 152, no. 1, pp. 107–151, 2022

    2022

  23. [23]

    Feature overview for joint modeling of sound event detection and localization using a microphone array,

    Daniel Krause, Archontis Politis, and Konrad Kowalczyk, “Feature overview for joint modeling of sound event detection and localization using a microphone array,” in2020 28th Euro- pean Signal Processing Conference (EUSIPCO). IEEE, 2021, pp. 31–35

    2021

  24. [24]

    Binamix – a python li- brary for generating binaural audio datasets,

    Dan Barry, Davoud Shariat Panah, Alessandro Ragano, Jan Skoglund, and Andrew Hines, “Binamix – a python li- brary for generating binaural audio datasets,”arXiv preprint arXiv:2505.01369, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    A perceptual evaluation of individual and non- individual hrtfs: A case study of the sadie ii database,

    Cal Armstrong, Lewis Thresh, Damian Murphy, and Gavin Kearney, “A perceptual evaluation of individual and non- individual hrtfs: A case study of the sadie ii database,”Applied Sciences, vol. 8, no. 11, pp. 2029, 2018

    2029

  26. [26]

    Tsp speech database,

    Peter Kabal, “Tsp speech database,”McGill University, Database V ersion, vol. 1, no. 0, pp. 09–02, 2002

    2002

  27. [27]

    Binaqual: A full-reference ob- jective localization similarity metric for binaural audio,

    Davoud Shariat Panah, Dan Barry, Alessandro Ragano, Jan Skoglund, and Andrew Hines, “Binaqual: A full-reference ob- jective localization similarity metric for binaural audio,”arXiv preprint arXiv:2505.11915, 2025

    arXiv 2025

  28. [28]

    The dominant role of low-frequency interaural time differences in sound localiza- tion,

    Frederic L Wightman and Doris J Kistler, “The dominant role of low-frequency interaural time differences in sound localiza- tion,”The Journal of the Acoustical Society of America, vol. 91, no. 3, pp. 1648–1661, 1992

    1992