Neural Directional Filtering with Configurable Directivity Pattern at Inference

Emanu\"el A. P. Habets; Srikanth Raj Chetupalli; Weilong Huang

REVIEW 3 major objections 5 minor 35 references

Neural Directional Filtering with Configurable Directivity Pattern at Inference

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A neural spatial filter whose directivity pattern is set by the user at run time: the paper claims a FiLM-conditioned non-linear filter on a four-microphone array can track any supplied azimuthal sensitivity pattern, outperforming an oracle

desk verdict A useful incremental extension — configurable directivity via FiLM — with solid anechoic simulations, but the 'user-defined' claim is narrower than advertised and the direct NDF baselines are missing. read the letter →

arxiv 2510.20253 v2 pith:JYOTK2YV submitted 2025-10-23 eess.AS

Weilong Huang , Srikanth Raj Chetupalli , Emanuël A. P. Habets This is my paper · ORCID

classification eess.AS

keywords neuraldirectionalfilteringdirectivitypatternmicrophonearrayspatialFiLMconditioningdeepnon-linearinference-timeconfigurabilitytrainingstrategies

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to establish that a compact microphone array can perform spatial filtering with a directivity pattern that the user selects on the spot, during inference, without changing the network weights or computing a new filter. The proposed method, UNDF, feeds a sampled version of the desired pattern into a deep non-linear filtering network using feature-wise linear modulation (FiLM), so the pattern acts as a conditioning signal rather than being fixed at training time. The central evidence is that this architecture generalizes to patterns never seen in training: higher differential orders, scaled gains, and new steering directions, whereas a simpler conditioning scheme that feeds the pattern only as an LSTM initial state fails on these unseen cases. A two-stage training recipe that mixes random differential-array patterns and rectangular masks through linear combinations is shown to further improve approximation, including irregular shapes. If correct, this replaces fixed beamformers and DOA-dependent parametric filters with one model whose sensitivity pattern can be swapped live, enabling dynamic audio zoom or target-adjustable capture.

What carries the argument

The key mechanism is a FiLM (feature-wise linear modulation) layer: the user pattern vector, sampled at 72 azimuth angles, is mapped through two linear layers to per-feature scale α and shift β (each of dimension 512), and every feature tensor from the preceding bidirectional LSTM is transformed as y = α ⊙ x + β, with the same modulation applied across all time frames and frequency bins. Inserted between the bidirectional and unidirectional LSTM layers of the joint non-linear filtering (JNF) architecture, this makes the pattern a global conditioning signal that the whole network can respond to. The second load-bearing component is the training recipe: patterns are generated not from a fixed

What would settle it

If a user-defined pattern far from the training family — say an extremely narrow mainlobe with sidelobes below the trained −20 dB floor, or a pattern that varies with frequency — produced large directivity approximation errors and an SDR drop relative to a simple delay-and-sum baseline in a controlled test, the generalization claim would be falsified.

Watch

Extended reading notes

Core claim

UNDF's claim is that a neural directional filter can be made configurable at inference: given a user-defined directivity pattern, expressed as a vector of gain values over azimuth angles, the network estimates a complex mask that, applied to a reference microphone signal, reproduces the target capture defined by that pattern. The decisive result in the paper is the comparison between the two conditioning mechanisms: FiLM-JNF, where the pattern vector modulates internal features via per-feature scaling and shifting, preserves the shape of unseen 3rd-order cardioid, scaled, and rotated patterns, whereas PV-JNF, which feeds the pattern only as an LSTM initial state, shows steering bias and poor

Load-bearing premise

The quantitative claims rest on an anechoic, co-planar, fixed-distance test setup where the training target is generated from the same pattern vector that is given to the network; outside this regime (reverberation, elevation, frequency-dependent patterns) the paper offers only a qualitative demonstration.

Editorial extensions

If this is right

Users can switch a compact microphone array's sensitivity pattern at run time, with no recomputation of beamforming weights and no explicit direction-of-arrival estimation.
Because the network is causal and conditioning is applied every frame, rapid changes in the pattern take effect immediately in the output stream, enabling time-varying directivity (demonstrated on a 20 s example).
The same FiLM-JNF model, trained only on speech in anechoic conditions, transfers qualitatively to music input and to a reverberant room.
Training with linearly combined random differential-array and rectangular patterns is what grants the model the ability to approximate irregular, non-DMA shapes; without the combination step, multi-scaled patterns are approximated poorly.
In simulated two-source anechoic scenes with unseen patterns, FiLM-JNF achieves higher SDR than a parametric filter that is handed the true source directions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An untested extension is frequency-dependent directivity patterns: because FiLM modulation is shared across all frequency bins, the current model is limited to frequency-invariant patterns; injecting a per-frequency pattern condition would be a natural next step.
The training recipe defines a pattern space, so the real capability boundary is how far the model stays reliable beyond that space — e.g., very narrow beams, patterns with multiple disjoint lobes, or elevation-dependent patterns — a quantitative pattern-coverage study would settle this.
The reverberant, time-varying demonstration is qualitative only; a systematic sweep over room RT60 and source distance would be needed to confirm the claimed robustness outside the anechoic training regime.
The same conditioning idea could be applied to other array geometries or distributed microphone nodes, where a user controls the filter by pattern rather than by a single target direction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The genuinely new thing here is that a trained neural directional filter can accept a user-supplied directivity pattern at inference as a conditioning vector, and in their anechoic simulations it generalizes to unseen steering angles, orders, and scaling factors. The FiLM conditioning is simple and sensible, and the training-recipe ablations are carefully done: Recipe B clearly beats Recipe A, and the comparison against the oracle-DOA parametric filter is a reasonable first baseline. The qualitative demo with time-varying patterns is a nice touch, and the paper is honest enough to state its main assumptions up front (zero elevation, frequency-invariant patterns, fixed distance).

That said, the central claim in the abstract — that users can define directivity patterns during inference — is broader than what the method actually supports. Because the FiLM parameters are shared across all time-frequency bins, the architecture cannot represent a frequency-dependent pattern at all. That is not a minor detail; it is a hard limitation of the conditioning mechanism. The paper only tests frequency-invariant patterns, so the scope gap is real, even though the assumptions are disclosed in Section 2. The one out-of-regime experiment (the reverberant demo) is qualitative only and uses time-block-constant patterns, so it does not close the gap.

A second soft spot is that the paper never compares against the authors' own fixed or steerable NDF methods [19, 20]. Those are the direct predecessors this work extends, and without them it is hard to tell whether the added flexibility costs anything in pattern fidelity or SDR. Third, all SDR numbers are point estimates with no error bars or significance tests; given the claims of improvement, this matters. And there is no release of code or weights, so the results are not independently reproducible.

None of this is fatal. The engineering idea is sound, the ablations are logical, and the paper is clearly written. It is a credible conference-level contribution that would benefit from a revision that: (1) explicitly limits all claims to frequency-invariant patterns, (2) adds the missing NDF baselines, and (3) reports variance or significance. I would send it to peer review for a venue like ICASSP or IWAENC, but I would ask for those revisions before acceptance.

Referee Report

3 major / 5 minor

Summary. The paper proposes UNDF, a neural spatial filter that accepts a user-defined directivity pattern vector as conditioning input at inference time. Two conditioning variants are compared: PV-JNF (pattern vector fed through a linear layer into a BiLSTM) and FiLM-JNF (feature-wise linear modulation inserted between BiLSTM and UniLSTM). In controlled anechoic simulations with co-planar, fixed-distance sources and frequency-invariant patterns, the FiLM-JNF variant is shown to generalize to unseen steering directions, higher DMA orders, and scaled patterns, and to achieve higher SDR than an oracle-DOA parametric filter. Training recipes A, B, B+, and B- are studied, with Recipe B+ improving approximation of irregular patterns. A qualitative reverberant-room demo illustrates time-varying pattern switching.

Significance. If the result holds, the paper offers a practical way to reconfigure a compact-array spatial filter at inference time without DOA estimation or filter recomputation, extending prior NDF work. The controlled simulation design is a strength: unseen patterns are genuinely held-out inputs, the parametric baseline is given oracle DOA information, and the architecture comparison (PV vs. FiLM) is informative. The paper does not provide code, and all quantitative evidence is obtained in a deliberately narrow regime; the abstract and conclusions currently overstate the breadth of what is demonstrated.

major comments (3)

[Abstract; §2; §3.1] The headline claim that UNDF generalizes to 'user-defined directivity patterns' and 'outperforms conventional methods' is not qualified in the Abstract or Conclusions. §2 restricts the study to anechoic, zero-elevation, fixed-distance (d=1.5 m) scenarios and frequency-invariant patterns; §3.1's FiLM modulation y=α⊙x+β shares α and β across all time-frequency bins, so frequency-dependent user patterns cannot in principle be represented. Eq. (2) weights only direct-path components at true source DOAs, so reverberation, varying range, and elevation are outside the training objective. Please reword the Abstract and Conclusions to state the demonstrated scope explicitly.
[§3.2, §5.2, Table 2] The training recipes are not fully specified. Recipe B says 'up to C random DMA patterns' but C is never defined, and Recipe B+ says 'linear combination of random rectangular patterns' and 'mixed' without giving counts or distributions. The test patterns 'multi-scaled pattern' and 'irregular pattern' in §5.2 and Table 2 are also not defined (no formula, angle values, or scaling factor). Without these definitions, the training-strategy comparison and the generalization results cannot be reproduced. Please provide C, the component counts/proportions, and the exact test-pattern definitions (or a code release).
[§5.1, Tables 1–2] The claim that FiLM-JNF has 'significantly better performance' is based on point estimates of SDR without any measure of variability or significance test. With 3240 test samples per condition, it should be possible to report standard deviations or confidence intervals (e.g., bootstrap over test samples). This is especially relevant in Table 2, where Recipe B+ (17.61 dB) is lower than Recipe B (19.75 dB) on the multi-scaled pattern; without uncertainty estimates, the relative merits of B versus B+ are not fully quantified.

minor comments (5)

[Abstract] Typo: 'during interference' should be 'during inference'.
[§3.1] The acronym 'PV-JNF' is not expanded; please define it (e.g., pattern-vector JNF).
[§5.3, Fig. 1] The text states that 'the model is causal [17,19]', but the FiLM-JNF architecture in Fig. 1 includes a BiLSTM, which is bidirectional and therefore non-causal over the input window. Please clarify how 'immediate' pattern switching is achieved, or process each constant-pattern segment separately and state so.
[§4.1] The sentence 'Maximum suppression for all patterns was set to −20 dB' is not explained. Specify how this is enforced (e.g., clipping the target pattern in dB, or as a loss term) so that the experimental setting is unambiguous.
[Fig. 3 caption] The caption says 'NDF models are trained based on the FiLM-JNF'; this should likely be 'UNDF models' for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UNDF target is independently defined by the acoustic model (Eq. 2) and evaluated against an external oracle-DOA baseline; self-citations are architectural context.

full rationale

The derivation chain is self-contained. The UNDF task is defined by Eq. (2), where the target signal is a physically meaningful pattern-weighted sum of source components at the reference microphone. The user-defined pattern vector is a conditioning input, but the network must learn the mapping from microphone signals to this target; the target is not a function of the network's fitted parameters. Generalization tests use new pattern vectors that were not in the training set (e.g., 3rd-order, scaled, and differently steered patterns), so these are genuine held-out predictions rather than re-statements of training targets. The main baseline is an external informed parametric filter with oracle DOA estimates, not a self-fit. Self-citations to [19] and [20] provide the backbone architecture and experimental configuration; they are incremental context and are not used to prove the central claim. The paper acknowledges its scope restrictions (anechoic, co-planar, fixed distance, frequency-invariant patterns) and labels the reverberant demo as qualitative, which are honest limitations rather than circular steps. Therefore no claim in the paper reduces by construction to its inputs, and no load-bearing self-citation chain is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-chosen training-pattern recipes, a synthetic anechoic direct-path data model, and the assumption that users specify frequency-invariant azimuth patterns. The architecture and conditioning are engineering constructions evaluated empirically; no new physical entities are postulated.

free parameters (4)

Training-pattern recipe distributions (A, B, B+) = J=1/μ grid 0-0.9/θs 0°-300° for A; random μ,θs,J∈[1,11] for B; B+ adds rectangular combos
These hand-designed training distributions determine which unseen patterns generalize; the B- vs B vs B+ ablation shows the results are sensitive to them.
Maximum suppression floor = -20 dB
Applied to all patterns in training/test; sets the achievable null depth and therefore affects every pattern-approximation and SDR number.
Pattern-vector sampling length L = 72
Discretization of the user-defined pattern; test source DOAs are offset from this grid, so off-grid interpolation is part of the central claim.
Training patterns per source-array setup P = 60
Together with 4320 training setups, determines the coverage of pattern space under each recipe.

assumptions (5)

domain assumption Acoustic scene is anechoic with zero-order reflections; signals are direct-path transfer functions.
Section 3.2 uses the RIR generator with reflection order zero. All quantitative results are obtained only in this regime.
domain assumption All sources are in the x-y plane with zero elevation and fixed distance d=1.5 m; the pattern depends only on azimuth.
Section 2 states that all incident sounds have zero elevation; Section 4.1 fixes source-array distance. Elevation-dependent patterns are outside the claim.
domain assumption The desired directivity pattern is frequency-invariant and exactly represented by L sampled values.
Section 2: 'we assume the pattern to be frequency-invariant... vector applicable to all frequency bands.' The method is not demonstrated for frequency-dependent patterns.
domain assumption The training target is the reference-microphone sum of source signals weighted by the pattern values (Eq. 2).
This defines the task and the supervision signal; it assumes user intent is correctly modeled as a linear weighting of source contributions.
domain assumption A single-channel complex mask applied to the reference microphone can realize the desired pattern.
The DNN outputs M[f,t] and estimates Z = M Y1; no theoretical guarantee is given that this mask restriction does not limit achievable patterns.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural Directional Filtering with Configurable Directivity Pattern at Inference." pith.science (2026). https://pith.science/paper/JYOTK2YV

@misc{pith2026251020253,
  author       = {Pith},
  title        = {Pith review of: Neural Directional Filtering with Configurable Directivity Pattern at Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYOTK2YV}},
  note         = {Machine review of arXiv:2510.20253}
}

read the original abstract

Spatial filtering with a desired directivity pattern is advantageous for many audio applications. In this work, we propose neural directional filtering with user-defined directivity patterns (UNDF), which enables spatial filtering based on directivity patterns that users can define during inference. To achieve this, we propose a DNN architecture that integrates feature-wise linear modulation (FiLM), allowing user-defined patterns to serve as conditioning inputs. Through analysis, we demonstrate that the FiLM-based architecture enables the UNDF to generalize to unseen user-defined patterns during interference with higher directivities, scaling variations, and different steering directions. Furthermore, we progressively refine training strategies to enhance pattern approximation and enable UNDF to approximate irregular shapes. Lastly, experimental comparisons show that UNDF outperforms conventional methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 2 linked inside Pith

[1]

INTRODUCTION Spatial filtering with a desired directivity pattern is beneficial to many audio applications. Traditional fixed beamforming [1–3] can achieve spatial filtering with a predefined directivity pattern by lin- early filtering microphone array signals, but it requires a large num- ber of microphones and a large array aperture not only to ensure a...
[2]

LetY q[f, t]represent the mix- ture signal at theq-th microphone in the short-time Fourier trans- form (STFT) domain, wherefandtdenote the frequency and time indices, respectively

PROBLEM FORMULATION We consider an acoustic scene in an anechoic environment, where Nsound sources are recorded by a compact array equipped with Qomnidirectional microphones. LetY q[f, t]represent the mix- ture signal at theq-th microphone in the short-time Fourier trans- form (STFT) domain, wherefandtdenote the frequency and time indices, respectively. T...

arXiv 2025
[3]

PROPOSED METHOD 3.1. DNN architecture and loss function We propose two DNN architectures (see Figure 1), which are adap- tations of the joint spatial and temporal-spectral non-linear filter- ing (FT-JNF) [17] for the UNDF task. For both architectures, the real and imaginary parts of theQmicrophone signals in the STFT domain are stacked along the channel d...
[4]

EXPERIMENTAL SETUP 4.1. Datasets and configurations We used speech signals from the ‘train-clean-360’ and ‘dev-clean’ subsets of the LibriSpeech corpus [24] as the source signals for train- ing and validation, respectively. For the test sets, speech samples were selected from the EARS dataset [25] with minimum loudness of−42dBFS [26] as the selection crit...
[5]

Conditioning methods We study and compare the two conditioning methods in FiLM-JNF and PV-JNF, using Recipe A, which consists only of a finite number of specific1 st-order patterns

RESULTS AND ANALYSIS 5.1. Conditioning methods We study and compare the two conditioning methods in FiLM-JNF and PV-JNF, using Recipe A, which consists only of a finite number of specific1 st-order patterns. Figure 2 (a) shows the pattern esti- mates for a1 st-order cardioid pattern steered towardsθ s = 60 ◦, 0° 30° 60° 90° 120° 150° 180° 210° 240° 270° 3...

arXiv 2026
[6]

For this, we integrated the FiLM layer into the JNF architecture, creating the FiLM-JNF model, which enables the conditioning of user-defined patterns

CONCLUSIONS In this work, we proposed a method for performing spatial filtering based on directivity patterns that users can define at inference, re- ferred to as UNDF. For this, we integrated the FiLM layer into the JNF architecture, creating the FiLM-JNF model, which enables the conditioning of user-defined patterns. Through analysis, we demon- strated ...
[7]

The hardware is funded by the German Research Foundation (DFG)

ACKNOWLEDGEMENTS The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Perfor- mance Computing Center (NHR@FAU) of the Friedrich-Alexander- Universit¨at Erlangen-N ¨urnberg (FAU). The hardware is funded by the German Research Foundation (DFG)
[8]

Superdirectional microphone arrays,

Gary W Elko, “Superdirectional microphone arrays,”Acoustic signal processing for telecommunication, pp. 181–237, 2000

2000

Show all 35 references

[9]

Michael Brandstein and Darren Ward,Microphone arrays: sig- nal processing techniques and applications, Springer Science & Business Media, 2001

2001
[10]

Fixed beamforming,

Jacob Benesty, Israel Cohen, and Jingdong Chen, “Fixed beamforming,”Fundamentals of Signal Enhancement and Ar- ray Signal Processing, pp. 237–282, 2018

2018
[11]

Spatial filtering using directional audio coding parameters,

Markus Kallinger, Giovanni Del Galdo, Fabian Kuech, Dirk Mahne, and Richard Schultz-Amling, “Spatial filtering using directional audio coding parameters,” inProc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP). IEEE, 2009, pp. 217–220

2009
[12]

Geometry-based spatial sound acqui- sition using distributed microphone arrays,

Oliver Thiergart, Giovanni Del Galdo, Maja Taseska, and Emanu¨el A. P. Habets, “Geometry-based spatial sound acqui- sition using distributed microphone arrays,”IEEE Trans. Aud., Sp., Lang. Proc., vol. 21, no. 12, pp. 2583–2594, 2013

2013
[13]

An informed LCMV filter based on multiple instantaneous direction-of- arrival estimates,

Oliver Thiergart and Emanu ¨el A. P. Habets, “An informed LCMV filter based on multiple instantaneous direction-of- arrival estimates,” inProc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP). IEEE, 2013, pp. 659–663

2013
[14]

An informed parametric spatial filter based on instantaneous direction-of-arrival estimates,

Oliver Thiergart, Maja Taseska, and Emanu ¨el A. P. Habets, “An informed parametric spatial filter based on instantaneous direction-of-arrival estimates,”IEEE Trans. Aud., Sp., Lang. Proc., vol. 22, no. 12, pp. 2182–2196, 2014

2014
[15]

Paramet- ric spatial sound processing: A flexible and efficient solution to sound scene acquisition, modification, and reproduction,

Konrad Kowalczyk, Oliver Thiergart, Maja Taseska, Giovanni Del Galdo, Ville Pulkki, and Emanu¨el A. P. Habets, “Paramet- ric spatial sound processing: A flexible and efficient solution to sound scene acquisition, modification, and reproduction,” IEEE Sig. Process. Mag., vol. 3...

2015
[16]

Sound field model violations in parametric spatial sound processing,

Oliver Thiergart and Emanu¨el A. P. Habets, “Sound field model violations in parametric spatial sound processing,” inProc. Intl. W. Ac. Sig. Enh. (IWAENC). VDE, 2012, pp. 1–4

2012
[17]

ReZero: Region-Customizable Sound Extraction,

Rongzhi Gu and Yi Luo, “ReZero: Region-Customizable Sound Extraction,”IEEE/ACM Trans. Aud., Sp., Lang. Proc., vol. 32, pp. 2576–2589, 2024

2024
[18]

The cone of silence: Speech sep- aration by localization,

Teerapat Jenrungrot, Vivek Jayaram, Steve Seitz, and Ira Kemelmacher-Shlizerman, “The cone of silence: Speech sep- aration by localization,” inProc. Advances in Neural Informa- tion Processing Systems (NeurIPS), 2020, vol. 33, pp. 20925– 20938

2020
[19]

Learning to separate voices by spatial regions,

Alan Xu and Romit Roy Choudhury, “Learning to separate voices by spatial regions,” inProc. Intl. Conf. on Machine Learning, 2022, vol. 162, pp. 24539–24549

2022
[20]

Deep audio zooming: Beamwidth-controllable neural beamformer,

Meng Yu and Dong Yu, “Deep audio zooming: Beamwidth-controllable neural beamformer,”arXiv preprint arXiv:2311.13075, 2023

2023 arXiv
[21]

Neural directed speech enhancement with dual micro- phone array in high noise scenario,

Wen Wen, Qiang Zhou, Yu Xi, Haoyu Li, Ziqi Gong, and Kai Yu, “Neural directed speech enhancement with dual micro- phone array in high noise scenario,” inProc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP). IEEE, 2025, pp. 1–5

2025
[22]

Nonlinear spatial filter- ing in multichannel speech enhancement,

Kristina Tesch and Timo Gerkmann, “Nonlinear spatial filter- ing in multichannel speech enhancement,”IEEE/ACM Trans. Aud., Sp., Lang. Proc., vol. 29, pp. 1795–1805, 2021

2021
[23]

Spatially selective deep non-linear filters for speaker extraction,

Kristina Tesch and Timo Gerkmann, “Spatially selective deep non-linear filters for speaker extraction,” inProc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP), 2023

2023
[24]

Insights into deep non- linear filters for improved multi-channel speech enhancement,

Kristina Tesch and Timo Gerkmann, “Insights into deep non- linear filters for improved multi-channel speech enhancement,” IEEE/ACM Trans. Aud., Sp., Lang. Proc., vol. 31, pp. 563–575, 2023

2023
[25]

Multi-channel speech separation using spatially selective deep non-linear filters,

Kristina Tesch and Timo Gerkmann, “Multi-channel speech separation using spatially selective deep non-linear filters,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 32, pp. 542–553, 2023

2023
[26]

Neural Directional Filtering: Far-field directivity control with a small microphone array,

Julian Wechsler, Srikanth Raj Chetupalli, Mhd Modar Hal- imeh, Oliver Thiergart, and Emanu ¨el A. P. Habets, “Neural Directional Filtering: Far-field directivity control with a small microphone array,” inProc. Intl. W. Ac. Sig. Enh. (IWAENC). IEEE, 2024, pp. 459–463

2024
[27]

Steerable neural directional filtering,

Weilong Huang, Mhd Modar Halimeh, Srikanth Raj Chetu- palli, Oliver Thiergart, and Emanu ¨el A. P. Habets, “Steerable neural directional filtering,” inProc. Forum Acusticum Eu- ronoise, European Acoustics Association, 2025

2025
[28]

FiLM: Visual reasoning with a general conditioning layer,

Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville, “FiLM: Visual reasoning with a general conditioning layer,” inProc. AAAI Conference on Artificial Intelligence, 2018, vol. 32

2018
[29]

John Eargle,The Microphone Book: From mono to stereo to surround-a guide to microphone design and application, Rout- ledge, 2012

2012
[30]

RIR generator,

Emanu ¨el A. P. Habets, “RIR generator,”https://github. com/ehabets/RIR-Generator, 2020, commit 3cf914d

2020
[31]

LibriSpeech: An ASR corpus based on public domain audio books,

Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “LibriSpeech: An ASR corpus based on public domain audio books,” inProc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP), 2015, pp. 5206–5210

2015
[32]

EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,

Julius Richter, Yi-Chiao Wu, Steven Krenn, Simon Welker, Bunlong Lay, Shinji Watanabe, Alexander Richard, and Timo Gerkmann, “EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,” in Proc. Interspeech, 2024, pp. 4873–4877

2024
[33]

Recommendation ITU-R BS.1770-5: Algorithms to measure audio programme loudness and true-peak audio level,

ITU-R, “Recommendation ITU-R BS.1770-5: Algorithms to measure audio programme loudness and true-peak audio level,” 2023

2023
[34]

Performance measurement in blind audio source separation,

Emmanuel Vincent, R ´emi Gribonval, and C ´edric F ´evotte, “Performance measurement in blind audio source separation,” IEEE/ACM Trans. Aud., Sp., Lang. Proc., vol. 14, no. 4, pp. 1462–1469, 2006

2006
[35]

SA-SDR: A novel loss function for separation of meeting style data,

Thilo von Neumann, Keisuke Kinoshita, Christoph Boeddeker, Marc Delcroix, and Reinhold Haeb-Umbach, “SA-SDR: A novel loss function for separation of meeting style data,” in Proc. IEEE Intl. Conf. on Ac., Sp. and Sig. Proc. (ICASSP). IEEE, 2022, pp. 6022–6026

2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

INTRODUCTION Spatial filtering with a desired directivity pattern is beneficial to many audio applications. Traditional fixed beamforming [1–3] can achieve spatial filtering with a predefined directivity pattern by lin- early filtering microphone array signals, but it requires a large num- ber of microphones and a large array aperture not only to ensure a...

[2] [2]

LetY q[f, t]represent the mix- ture signal at theq-th microphone in the short-time Fourier trans- form (STFT) domain, wherefandtdenote the frequency and time indices, respectively

PROBLEM FORMULATION We consider an acoustic scene in an anechoic environment, where Nsound sources are recorded by a compact array equipped with Qomnidirectional microphones. LetY q[f, t]represent the mix- ture signal at theq-th microphone in the short-time Fourier trans- form (STFT) domain, wherefandtdenote the frequency and time indices, respectively. T...

arXiv 2025

[3] [3]

PROPOSED METHOD 3.1. DNN architecture and loss function We propose two DNN architectures (see Figure 1), which are adap- tations of the joint spatial and temporal-spectral non-linear filter- ing (FT-JNF) [17] for the UNDF task. For both architectures, the real and imaginary parts of theQmicrophone signals in the STFT domain are stacked along the channel d...

[4] [4]

EXPERIMENTAL SETUP 4.1. Datasets and configurations We used speech signals from the ‘train-clean-360’ and ‘dev-clean’ subsets of the LibriSpeech corpus [24] as the source signals for train- ing and validation, respectively. For the test sets, speech samples were selected from the EARS dataset [25] with minimum loudness of−42dBFS [26] as the selection crit...

[5] [5]

Conditioning methods We study and compare the two conditioning methods in FiLM-JNF and PV-JNF, using Recipe A, which consists only of a finite number of specific1 st-order patterns

RESULTS AND ANALYSIS 5.1. Conditioning methods We study and compare the two conditioning methods in FiLM-JNF and PV-JNF, using Recipe A, which consists only of a finite number of specific1 st-order patterns. Figure 2 (a) shows the pattern esti- mates for a1 st-order cardioid pattern steered towardsθ s = 60 ◦, 0° 30° 60° 90° 120° 150° 180° 210° 240° 270° 3...

arXiv 2026

[6] [6]

For this, we integrated the FiLM layer into the JNF architecture, creating the FiLM-JNF model, which enables the conditioning of user-defined patterns

CONCLUSIONS In this work, we proposed a method for performing spatial filtering based on directivity patterns that users can define at inference, re- ferred to as UNDF. For this, we integrated the FiLM layer into the JNF architecture, creating the FiLM-JNF model, which enables the conditioning of user-defined patterns. Through analysis, we demon- strated ...

[7] [7]

The hardware is funded by the German Research Foundation (DFG)

ACKNOWLEDGEMENTS The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Perfor- mance Computing Center (NHR@FAU) of the Friedrich-Alexander- Universit¨at Erlangen-N ¨urnberg (FAU). The hardware is funded by the German Research Foundation (DFG)

[8] [8]

Superdirectional microphone arrays,

Gary W Elko, “Superdirectional microphone arrays,”Acoustic signal processing for telecommunication, pp. 181–237, 2000

2000