Pith. sign in

REVIEW 5 major objections 7 minor 83 references

Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and Hearing Loss Compensation

T0 review · 5 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read A single hearing-aid network that fuses the listener's audiogram into the denoiser via affine modulation outperforms serial and in-context-fusion designs on HASQI and PESQ at low computational cost.

desk verdict Credible design and honest ablations, but the SOTA comparison rests on unverified baseline conditioning and an undocumented audiogram split. read the letter →

arxiv 2509.07341 v1 pith:PW5Y2ZIZ submitted 2025-09-09 eess.AS

classification eess.AS
keywords audiogramfusionhearinglosscompensationnoisereductionpersonalizedspeechenhancementaffinemodulationaidsConformervoiceactivitydetection
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

The paper argues that the two jobs a hearing aid does in noise—removing background noise and compensating for the wearer's hearing loss—should be solved by one network explicitly conditioned on the patient's audiogram, not by running separate algorithms one after the other. Its central claim is that how the audiogram enters the network matters: repeating or naively concatenating the six sparse clinical thresholds with the spectrogram is inferior to linearly interpolating those thresholds across the full frequency axis and then injecting the result through affine modulation (learned scale, shift, and gate parameters) inside frequency- and temporal-wise Conformer blocks. The paper also adds a voice-activity-detection side task so the compensation gain does not amplify residual noise in silent gaps, a known failure of wide-dynamic-range-compression fitting. On synthetic DNS-Challenge and LibriSpeech+Demand test sets, the resulting model reports the best HASQI and PESQ scores among the compared joint models while using a fraction of the compute of larger baselines. If the claim holds, hearing-aid processing can be systematically optimized as one learned task instead of stacked hand-tuned stages.

What carries the argument

The load-bearing object is the AMFT-Conformer, the affine modulation-based audiogram fusion frequency-temporal Conformer. A SiLU-plus-linear transform maps the encoded audiogram into two groups of affine parameters (gamma, beta, alpha) per fusion stage; one group re-scales and shifts the frequency-oriented spectrum features before the Conformer block, the second after it, with a residual gated connection—then the same procedure repeats along the time axis. This is the mechanism that turns a static patient profile into adaptive, per-frequency and per-time conditioning of the network's internal representation. Its companion is HL-Linear, the linear interpolation of the six measured thresholds

What would settle it

Two checks would settle the central claim. First, re-run the DNS-Challenge and LibriSpeech+Demand comparisons with every baseline's audiogram-input path specified and equalized (for instance, the same HL-Encoder output concatenated into each architecture), and see whether AFN-HearNet still leads on HASQI and PESQ; the paper leaves this unspecified. Second, run a blinded listening test with hearing-impaired listeners on real recorded noisy speech and real audiograms, since the reported gains are computed with objective metrics on synthetic mixtures.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that joint noise reduction and hearing-loss compensation improve when the audiogram—a static profile of six clinical thresholds—is linearly interpolated to the spectrogram's frequency resolution, encoded by a small ConvMLP, and fused into the speech representation repeatedly through affine modulation: learned scale, shift, and gate parameters applied before and after each frequency- and temporal-wise Conformer block, a mechanism the paper calls a neural analogue of auditory modulation filtering. A VAD auxiliary head embeds speech/non-speech awareness, protecting silent segments from excess compensation gain. Ablations on FT-LSTM and FT-Conformer backbo

Load-bearing premise

The headline comparison treats CRN, DCCRN, CMGAN, NUNet_TLS, and CompNet as 'in-context fusion joint models,' but none of these architectures natively accepts an audiogram and the paper never specifies how they were modified to receive one; if those baselines were not given a real audiogram-conditioning mechanism, the reported superiority is not a fair test of fusion strategy.

Editorial extensions

If this is right

  • Serial NR-then-compression hearing-aid topologies could be replaced by one jointly trained network, eliminating the failure mode where wide-dynamic-range compression amplifies residual noise in silent gaps.
  • The affine-modulation fusion recipe transfers across backbone sequence models (both LSTM and Conformer baselines gain over in-context fusion), suggesting it is a general way to inject static patient profiles into time-frequency enhancement networks.
  • At 1.28M parameters, 3.39 GFLOPs, and a theoretical 16 ms latency, the reported gains come at a size compatible with real-time wearable deployment.
  • The double-mismatch generalization analysis (unseen speech and noise corpora) indicates the fused-plus-VAD model retains more of its performance under domain shift, which matters for real-world hearing-aid use.
  • Because HASQI is non-differentiable, the MetricGAN-based adversarial loss is the load-bearing piece that lets the network optimize the hearing-aid quality metric directly rather than a proxy.

Reading between the lines

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

  • My reading, extending the paper: the audiogram-as-conditioning idea is an instance of a broader principle—sparse, static clinical measurements can steer a deep enhancer through affine modulation—which could transfer to other assistive devices such as cochlear-implant front ends or personalized beamforming without architectural redesign.
  • The paper's own results suggest that explicit VAD supervision becomes partly redundant once the backbone has global attention, so the natural next step is auxiliary tasks encoding information attention cannot self-learn, such as noise type or SNR class.
  • An audiogram-conditioned generative model is the untested competitor implied by the SGMSE comparison: the two-stage diffusion-plus-compression baseline loses on PESQ despite higher SDR, so conditioning a generative model directly on the audiogram would test whether joint optimization or the discriminative backbone is responsible for the perceptual gain.
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.

Referee Report

5 major / 7 minor

Summary. The paper proposes AFN-HearNet, a joint noise reduction (NR) and hearing loss compensation (HLC) model for hearing aids. The core ideas are (i) an audiogram-specific encoder that linearly interpolates the sparse clinical audiogram (6 thresholds) to the STFT frequency resolution before encoding, (ii) an affine modulation-based fusion mechanism (AMFT-Conformer) that uses the audiogram representation to scale, shift, and gate deep spectral features in both frequency and temporal Conformer blocks, and (iii) a VAD auxiliary training task intended to make the learned representation aware of speech/non-speech segments. Experiments are conducted on two synthetic datasets (DNS-Challenge and LibriSpeech+Demand) with audiograms from NHANES, using HASQI, PESQ, SDR, SI-SNR, and STOI as metrics. Ablations on FT-LSTM and FT-Conformer backbones, statistical t-tests, complexity analysis, and comparisons with several speech-enhancement/hearing-aid baselines are reported. The headline claim is that AFN-HearNet significantly outperforms state-of-the-art 'in-context fusion' joint models on HASQI and PESQ while retaining low computational cost.

Significance. If the comparison were fully supported, the paper would make a useful contribution to personalized speech enhancement: the linear-interpolation audiogram encoder is a simple, well-motivated idea; the affine modulation fusion is a plausible alternative to concatenation; and the efficiency results (1.28M parameters, 3.39 GFLOPs) are attractive for hearing-aid deployment. The manuscript has notable strengths: extensive ablations on two backbones, two datasets, t-tests for many comparisons, honest discussion of the modest VAD gains on the Conformer backbone, and a stated intent to release code. However, the central SOTA claim depends on a baseline comparison whose audiogram-conditioning is never described, and the audiogram train/test split is not documented. These gaps make the headline result unverifiable as written. The internal ablations (e.g., HA-MGAN vs. +AMFT-LSTM) provide some evidence for the affine-modulation contribution independent of the problematic external baselines, but the cross-model SOTA claim needs substantial clarification.

major comments (5)
  1. [Section 5.1, Algorithm 1] The manuscript labels CRN, DCCRN, CMGAN, NUNet_TLS, CompNet, DB-AIAT, MP-SENet, FSPEN, SEMamba, and PrimeK-Net as 'jointly address the NR and HLC tasks using the in-context fusion strategy.' None of these architectures in their published form takes an audiogram as input. The paper never specifies how each baseline was modified to accept audiogram information, which audiogram-extension method was used, whether they were retrained on the joint target (compensated speech), or whether the same audiogram sampling was applied. Without these details, the reader cannot verify that the comparisons in Tables 5 and 8 are joint NR+HLC systems; if these rows are standard speech-enhancement models, the t-tests in Tables 6 and 9 only show that audiogram-aware compensation improves metrics, not that affine modulation outperforms in-context fusion. This is load-bearing for the central claim and must be f
  2. [Section 6.1.3, Table 2] Algorithm 1 samples an audiogram uniformly at random for each training sample (line 19), but the paper does not document any train/test split of the 114 NHANES audiograms. If test samples use audiograms that appeared in training, the reported HASQI/PESQ may reflect memorization of specific hearing profiles rather than generalization to unseen listeners. The paper should report how the audiograms were partitioned (e.g., by participant or ear) and confirm that the test set uses held-out audiograms. This is particularly important because the model is personalized to the audiogram, and the claim of generalization across hearing-loss profiles depends on it.
  3. [Section 5.2, Eq. (14)] The VAD auxiliary module is presented as a contribution, and the third highlight states that 'Voice activation probability information contributes to personalized speech enhancement.' However, adding VAD to the full model (+AMFT-Conformer & HL-Encoder & VAD) decreases SDR and SI-SNR on both validation and test sets relative to +AMFT-Conformer & HL-Encoder: test SDR drops from 13.92 to 13.86, SI-SNR from 13.68 to 13.62; validation SDR drops from 13.86 to 13.75. The text acknowledges 'more modest' gains but does not mention these decreases. Moreover, Table 4 tests the full model against the baseline, not the incremental contribution of VAD over the +AMFT-Conformer & HL-Encoder configuration. For a claimed contribution, the incremental effect should be tested and discussed; as written, the evidence is mixed.
  4. [Section 6.1.4, Fig. 7] The implementation details state that hyperparameters α, λ, and μ are set for the discriminator, PMSQE, and multi-resolution STFT loss, respectively. However, Eq. (14) contains α, λ, and a fixed factor 0.3 for the focal loss, but no μ. Either μ is missing from the loss equation or it is unused; this should be clarified to make the loss function reproducible.
  5. [Section 6.1.3] The generalization-gap analysis in Fig. 7 is difficult to parse: the axis labels are garbled, the legend is missing, and the definition of ΔE_ref in the generation-gap formula (GE = ΔE − ΔE_ref / ΔE_ref) is not explicitly stated in the text. Since the section draws conclusions about the VAD module's cross-condition robustness, the figure and the associated definition need to be made interpretable.
minor comments (7)
  1. [Fig. 13 caption] Typos: 'AFA-HearNet-base' in Table 2 should be 'AFN-HearNet-base'; 'HL-Liner' in Table 3 should be 'HL-Linear'; 'ACFT-Conforme' in §6.1.3 should be 'AMFT-Conformer'.
  2. [Section 4.6] Panel (h) is labeled 'ACC-HearNet' but the model is AFN-HearNet.
  3. [Table 4] The term 'in-context fusion' is used for the concatenation baseline in Section 2.2 and for the baselines in Section 5.3, but the distinction between 'in-context' and 'affine modulation' could be defined explicitly at first use, as the reader may otherwise confuse it with in-context learning.
  4. [Section 4.3, Eq. (8)] The p-values are reported as '0.00+' and '0.05' etc. It would be clearer to use a consistent notation such as 'p<0.001' for values that round to 0.00.
  5. [Section 3, Eq. (5)] Equation (8) uses δ_i, β_i, and rearrange operations that are not all defined at the point of use; the subscript/superscript conventions for the two convolution layers are also inconsistent (the text says 'i∈[0,1]' but the equation shows two conv blocks). Please clean up the notation.
  6. [Section 6.2.1] In Eq. (5), the joint masking parameter is written as M(l,k)·Ḡ(l,b). The bar accent on G is defined only later; a brief definition immediately after the equation would help.
  7. [Appendix A] The text says 'AFN-HearNet achieves the highest scores across key speech quality metrics' but NUNet_TLS has higher SDR and SI-SNR (13.98/13.71 vs 13.86/13.62). Please qualify the statement to 'highest among the compared joint models' and acknowledge the SDR/SI-SNR trade-off explicitly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are empirical, evaluated on held-out synthetic data; the only self-citation (HA-MGAN baseline) is not load-bearing in a definitional sense.

full rationale

AFN-HearNet does not claim a first-principles derivation; its contributions are architectural (HL-Linear audiogram interpolation, affine-modulation fusion, VAD auxiliary task) and are assessed on held-out synthetic test sets. The core claim that affine modulation outperforms in-context fusion is an empirical ablation against HA-MGAN [8], the authors' own prior model, and against general speech-enhancement baselines; beating a self-built baseline is not a definitional reduction. HASQI appears both as a MetricGAN training target (Eq. 14) and as a headline evaluation metric, but the test-set HASQI is computed on data not used for training, so it remains a genuine prediction rather than a fitted value. Concerns that the baselines' audiogram conditioning is unspecified (Section 5.3, Tables 5/8) and that hyperparameters are selected after inspecting test-set results (Section 6.1.5, Fig. 8) are experimental-validity issues, not circularity: no equation, parameter, or model output is defined in terms of the claimed result. The self-citations [8,41,77] provide the baseline and audiogram feature-extension method, but they do not carry the derivation. No circular step can be exhibited under the required standard.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard supervised learning assumptions plus domain-specific premises that are stated but not derived: smoothness of audiometric thresholds, validity of FIG6-based synthetic targets, objectivity of HASQI/PESQ as hearing-aid proxies, representativeness of synthetic noisy mixtures, and an implicit disjoint audiogram split. No new physical entities are introduced.

free parameters (6)
  • adversarial loss weight alpha = 0.5 (tested range 0.5 to 2.5)
    Chosen via validation experiments in Appendix A; balances HASQI-related adversarial loss against other terms.
  • PMSQE loss weight lambda = 0.3 (tested 0.30 to 2.00)
    Selected on validation; larger lambda improves PESQ but slightly reduces STOI.
  • MR-STFT loss weight mu = 1
    Fixed in Eq. (14); no sensitivity analysis reported.
  • VAD focal loss weight = 0.3
    Set in Eq. (14); no ablation on this weight is reported.
  • middle channels C = 48 (C48DF4)
    Selected from C in {36,48,60} and downsampling factor in {4,8} based on the performance-complexity trade-off in Fig. 8.
  • number of AMFT-Conformer modules = 2
    Set in Section 5.2; no ablation is shown for this hyperparameter.
assumptions (5)
  • domain assumption Human hearing sensitivity varies smoothly between standard audiometric test frequencies, justifying linear interpolation.
    Invoked in Sections 1 and 4.3 to motivate HL-Linear; grounded in Fletcher-Munson contours [19] and the CASP auditory model [32].
  • domain assumption FIG6-based WDRC output for speech-in-quiet is an appropriate ground-truth target for hearing-loss compensation.
    Used to synthesize targets in Algorithm 1; the paper modifies non-speech segments, so the target is a hybrid of WDRC output and unprocessed speech.
  • domain assumption HASQI and PESQ computed on the synthetic target reflect real hearing-aid benefit.
    All evaluation and part of the loss use these objective metrics; no listening test with hearing-impaired listeners is conducted.
  • domain assumption Synthetic noisy mixtures with additive noise and gain perturbations generalize to real hearing-aid acoustic conditions.
    Datasets are generated by Algorithm 1 from clean speech and noise corpora; no real recordings or hearing-aid hardware are used.
  • domain assumption The 114 NHANES audiograms are split disjointly between training and test sets.
    The split is not described; if audiograms overlap between training and test, reported improvements could be inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and Hearing Loss Compensation." pith.science (2026). https://pith.science/paper/PW5Y2ZIZ

@misc{pith2026250907341,
  author       = {Pith},
  title        = {Pith review of: Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and Hearing Loss Compensation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PW5Y2ZIZ}},
  note         = {Machine review of arXiv:2509.07341}
}
read the original abstract

Hearing aids (HAs) are widely used to provide personalized speech enhancement (PSE) services, improving the quality of life for individuals with hearing loss. However, HA performance significantly declines in noisy environments as it treats noise reduction (NR) and hearing loss compensation (HLC) as separate tasks. This separation leads to a lack of systematic optimization, overlooking the interactions between these two critical tasks, and increases the system complexity. To address these challenges, we propose a novel audiogram fusion network, named AFN-HearNet, which simultaneously tackles the NR and HLC tasks by fusing cross-domain audiogram and spectrum features. We propose an audiogram-specific encoder that transforms the sparse audiogram profile into a deep representation, addressing the alignment problem of cross-domain features prior to fusion. To incorporate the interactions between NR and HLC tasks, we propose the affine modulation-based audiogram fusion frequency-temporal Conformer that adaptively fuses these two features into a unified deep representation for speech reconstruction. Furthermore, we introduce a voice activity detection auxiliary training task to embed speech and non-speech patterns into the unified deep representation implicitly. We conduct comprehensive experiments across multiple datasets to validate the effectiveness of each proposed module. The results indicate that the AFN-HearNet significantly outperforms state-of-the-art in-context fusion joint models regarding key metrics such as HASQI and PESQ, achieving a considerable trade-off between performance and efficiency. The source code and data will be released at https://github.com/deepnetni/AFN-HearNet.

Figures

Figures reproduced from arXiv: 2509.07341 by the authors.

Figure 1
Figure 1. Examples of NR and HLC processes in hearing aids include: serial concatenation, such as (a) preceding NR and (b) preceding HLC, each followed by another component; and (c) parallel NR and HLC, where both NR and the WDRC algorithm are applied independently to the noisy speech. silent region with negligible noise silent region with negligible noise [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Example of a compensated speech spectrogram, where dashed boxes indicate silent regions with negligible noise. function of outer hair cells in the cochlea by providing level-dependent amplification, thereby improving audibility while maintaining loudness comfort. Despite being the most effective means of hearing rehabilitation, HLC inevitably amplifies not only speech but also background noise, which substantially d… view at source ↗
Figure 3
Figure 3. The overall architecture of the AFN-HearNet model. optimizing either the NR or the HLC component independently, or sequentially combining the two. For the NR task, prior studies [63, 66, 39, 74] explore the potential of the DNN-based NR components for HAs, where filter coefficients are estimated by a DNN model based on input features. Regarding the HLC task, studies [15, 38] propose neural networks for compensating … view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Detailed architecture of the audiogram encoder module. insights, we introduce affine modulation into time–frequency modeling and further propose a post-gating strategy to emulate auditory masking effects, thereby enabling more effective fusion of audiogram and spectrum…
Figure 5
Figure 5. Figure 5: The differences between existing in-context and our proposed affine modulation audiogram fusion strategy. SiLU Linear Frequency-wise Conformer Block Affine Modulation Affine Modulation MLP SiLU Linear Temporal-wise Conformer Block Affine Modulation Affine Modulation ML…
Figure 6
Figure 6. Figure 6: Details architecture of the modulation fusion AMFT-Conformer module. a multi-layer convolutional encoder, a common fusion method for combining features from multiple models [27, 25]. However, since the fusion stage relies entirely on convolution operations and is proce…
Figure 7
Figure 7. Figure 7: Results of the ablation study variants under matched and mismatched acoustic conditions. DNS–DNS and Libri–Libri correspond to matched settings, where training and testing are conducted on the DNS-Challenge dataset and the LibriSpeech+DEMAND dataset, respectively. DNS–…
Figure 8
Figure 8. Figure 8: The impact of different configurations on model performance and complexity, with respect to the number of middle channels and the frequency dimension down-sampling factor. Each line in panels (a)-(c) comprises three points representing C36, C48, and C60, respectively, …
Figure 9
Figure 9. Figure 9: Proportional contribution of MACs from different network modules to the total computational cost. These findings suggest that although the VAD auxiliary training task does not yield substantial improvements for the Conformer backbone under matched conditions, it nevert…
Figure 10
Figure 10. Figure 10: Performance–efficiency trade-offs for different models on the DNS-Challenge test set, showing the relationship between GFLOPs and HASQI, WB-PESQ, and SDR. The red dashed line indicates the Pareto frontier. 0. 64 0. 75 0. 78 0. 76 0.80 0. 75 0.80 0.82 0.83 1 .32 2. 01 …
Figure 11
Figure 11. Figure 11: Metric results from various joint models on the DNS-Challenge test set. The median is indicated by empty diamonds within the violin plots [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Performance–efficiency trade-offs for different baselines on the LibriSpeech+Demand test set. (a) Noisy Noisy Clean (b) Target Target Clean (c) DB-AIAT Target DB-AIAT (d) MP-SENet Target MP-SENet (e) FSPEN Target FSPEN (f) SEMamba Target SEMamba (g) PrimeK-Net Target …
Figure 13
Figure 13. Figure 13: Comparison of waveform and spectrum enhancement results from various joint models on the Lib￾riSpeech+Demand test set. In contrast, the jointly optimized AFN-HearNet demonstrates a clear advantage in speech quality, indicating that integrating enhancement and compensa…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

83 extracted references · 75 canonical work pages

  1. [1]

    Speech enhancement using an adaptive wiener filtering approach

    Abd El-Fattah, M., Dessouky, M.I., Diab, S., Abd El-Samie, F., 2008. Speech enhancement using an adaptive wiener filtering approach. Progress In Electromagnetics Research M 4, 167–184

  2. [2]

    Abd El-Fattah, M.A., Dessouky, M.I., Abbas, A.M., Diab, S.M., El-Rabaie, E.S.M., Al-Nuaimy, W., Alshebeili, S.A., Abd El-samie, F.E.,

  3. [3]

    Akeroyd, M.A., Bailey, W., Barker, J., Cox, T.J., Culling, J.F., Graetzer, S., Naylor, G., Podwińska, Z., Tu, Z., 2023. The 2nd clarity enhancementchallengeforhearingaidspeechintelligibilityenhancement:Overviewandoutcomes,in:ICASSP2023-2023IEEEInternational Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1–5

  4. [4]

    Classification of hearing loss

    Alshuaib, W.B., Al-Kandari, J.M., Hasan, S.M., 2015. Classification of hearing loss. Update on hearing loss 4, 29–37

  5. [5]

    Cmgan: Conformer-based metric gan for speech enhancement

    Cao, R., Abdulatif, S., Yang, B., 2022. Cmgan: Conformer-based metric gan for speech enhancement. arXiv preprint arXiv:2203.15149

  6. [6]

    Semantic segmentation with second-order pooling, in: European conference on computer vision, Springer

    Carreira, J., Caseiro, R., Batista, J., Sminchisescu, C., 2012. Semantic segmentation with second-order pooling, in: European conference on computer vision, Springer. pp. 430–443

  7. [7]

    An investigation of incorporating mamba for speech enhancement

    Chao, R., Cheng, W.H., La Quatra, M., Siniscalchi, S.M., Yang, C.H.H., Fu, S.W., Tsao, Y., 2024. An investigation of incorporating mamba for speech enhancement. arXiv preprint arXiv:2405.06573

  8. [8]

    Speech denoising and compensation for hearing aids using an ftcrn-based metric gan

    Cheng, J., Liang, R., Zhao, L., Huang, C., Schuller, B.W., 2023. Speech denoising and compensation for hearing aids using an ftcrn-based metric gan. IEEE Signal Processing Letters 30, 374–378

Show all 83 references
  1. [9]

    Residual fusion probabilistic knowledge distillation for speech enhancement

    Cheng, J., Liang, R., Zhou, L., Zhao, L., Huang, C., Schuller, B.W., 2024. Residual fusion probabilistic knowledge distillation for speech enhancement. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32, 2680–2691

  2. [10]

    Modeling the effects of dynamic range compression on signals in noise

    Corey, R.M., Singer, A.C., 2021. Modeling the effects of dynamic range compression on signals in noise. The Journal of the Acoustical Society of America 150, 159–170

  3. [11]

    Cornell, S., Wang, Z.Q., Masuyama, Y., Watanabe, S., Pariente, M., Ono, N., Squartini, S., 2023. Multi-channel speaker extraction with adversarialtraining:Thewavlabsubmissiontotheclarityicassp2023grandchallenge,in:ICASSP2023-2023IEEEInternationalConference on Acoustics, Speech...

  4. [12]

    Attentionalfeaturefusion,in:ProceedingsoftheIEEE/CVFwinterconference on applications of computer vision, pp

    Dai,Y.,Gieseke,F.,Oehmcke,S.,Wu,Y.,Barnard,K.,2021. Attentionalfeaturefusion,in:ProceedingsoftheIEEE/CVFwinterconference on applications of computer vision, pp. 3560–3569

  5. [13]

    Modeling auditory processing of amplitude modulation

    Dau, T., Kollmeier, B., Kohlrausch, A., 1997. Modeling auditory processing of amplitude modulation. i. detection and masking with narrow- band carriers. The Journal of the Acoustical Society of America 102, 2892–2905

  6. [14]

    Real time speech enhancement in the waveform domain

    Defossez, A., Synnaeve, G., Adi, Y., 2020. Real time speech enhancement in the waveform domain. arXiv preprint arXiv:2006.12847

  7. [15]

    Dynamic processing neural network architecture for hearing loss compensation

    Drgas, S., Bramsløw, L., Politis, A., Naithani, G., Virtanen, T., 2023. Dynamic processing neural network architecture for hearing loss compensation. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32, 203–214

  8. [16]

    Gated recurrent fusion with joint training framework for robust end-to-end speech recognition

    Fan, C., Yi, J., Tao, J., Tian, Z., Liu, B., Wen, Z., 2020. Gated recurrent fusion with joint training framework for robust end-to-end speech recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 29, 198–209

  9. [17]

    Compnet: Complementary network for single-channel speech enhancement

    Fan, C., Zhang, H., Li, A., Xiang, W., Zheng, C., Lv, Z., Wu, X., 2023. Compnet: Complementary network for single-channel speech enhancement. Neural Networks 168, 508–517

  10. [18]

    Hearingaidsformildtomoderatehearing loss in adults

    Ferguson,M.A.,Kitterick,P.T.,Chong,L.Y.,Edmondson-Jones,M.,Barker,F.,Hoare,D.J.,2017. Hearingaidsformildtomoderatehearing loss in adults. Cochrane Database of Systematic Reviews

  11. [19]

    Loudness, its definition, measurement and calculation

    Fletcher, H., Munson, W.A., 1933. Loudness, its definition, measurement and calculation. Bell System Technical Journal 12, 377–430

  12. [20]

    Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement, in: International Conference on Machine Learning, PMLR

    Fu, S.W., Liao, C.F., Tsao, Y., Lin, S.D., 2019. Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement, in: International Conference on Machine Learning, PMLR. pp. 2031–2041

  13. [21]

    Compactbilinearpooling,in:ProceedingsoftheIEEEconferenceoncomputervisionand pattern recognition, pp

    Gao,Y.,Beijbom,O.,Zhang,N.,Darrell,T.,2016. Compactbilinearpooling,in:ProceedingsoftheIEEEconferenceoncomputervisionand pattern recognition, pp. 317–326

  14. [22]

    Assessing the generalization gap of learning-based speech enhancement systems in noisy and reverberant environments

    Gonzalez, P., Alstrøm, T.S., May, T., 2023. Assessing the generalization gap of learning-based speech enhancement systems in noisy and reverberant environments. IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, 3390–3403

  15. [23]

    Controllable joint noise reduction and hearing loss compensation using a differentiable auditory model

    Gonzalez, P., Dau, T., May, T., 2025. Controllable joint noise reduction and hearing loss compensation using a differentiable auditory model. arXiv preprint arXiv:2507.09372

  16. [24]

    Graetzer, S., Barker, J., Cox, T.J., Akeroyd, M., Culling, J.F., Naylor, G., Porter, E., Viveros Munoz, R., 2021. Clarity-2021 challenges: Machine learning challenges for advancing hearing aid processing, in: Proceedings of the Annual Conference of the International Speech Com...

  17. [25]

    Mapfusion: A novel bev feature fusion network for multi-modal map construction

    Hao, X., Diao, Y., Wei, M., Yang, Y., Hao, P., Yin, R., Zhang, H., Li, W., Zhao, S., Liu, Y., 2025. Mapfusion: A novel bev feature fusion network for multi-modal map construction. arXiv preprint arXiv:2502.04377

  18. [26]

    Neural speech enhancement with unsupervised pre-training and mixture training

    Hao, X., Xu, C., Xie, L., 2023. Neural speech enhancement with unsupervised pre-training and mixture training. Neural Networks 158, 216–227. Ni et al.:Preprint submitted to Information Fusion Page 22 of 27 Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduct...

  19. [27]

    Mbfusion: A new multi-modal bev feature fusion method for hd map construction, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE

    Hao, X., Zhang, H., Yang, Y., Zhou, Y., Jung, S., Park, S.I., Yoo, B., 2024. Mbfusion: A new multi-modal bev feature fusion method for hd map construction, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE. pp. 15922–15928

  20. [28]

    Distributed energy-saving speech enhancement in wireless acoustic sensor networks

    Hu, D., Si, Q., Bao, F., Zhang, H., 2025. Distributed energy-saving speech enhancement in wireless acoustic sensor networks. Information Fusion 113, 102593

  21. [29]

    Dccrn: Deepcomplex convolutionrecurrent networkfor phase-aware speech enhancement

    Hu, Y.,Liu, Y.,Lv, S.,Xing, M.,Zhang, S., Fu,Y., Wu,J., Zhang,B., Xie, L.,2020. Dccrn: Deepcomplex convolutionrecurrent networkfor phase-aware speech enhancement. arXiv preprint arXiv:2008.00264

  22. [30]

    Hwang,S.,Park,Y.,Park,S.,2022.Monoauralspeechenhancementusinganestedu-netwithtwo-levelskipconnections.,in:INTERSPEECH, pp. 191–195

  23. [31]

    Weibull and nakagami speech priors based regularized nmf with adaptive wiener filter for speech enhancement

    Jannu, C., Vanambathina, S.D., 2023. Weibull and nakagami speech priors based regularized nmf with adaptive wiener filter for speech enhancement. International Journal of Speech Technology 26, 197–209

  24. [32]

    A computational model of human auditory signal processing and perception

    Jepsen, M.L., Ewert, S.D., Dau, T., 2008. A computational model of human auditory signal processing and perception. The Journal of the Acoustical Society of America 124, 422–438

  25. [33]

    The 3 types of sensorineural hearing loss: Loudness and intelligibility considerations

    Killion, M.C., 1993. The 3 types of sensorineural hearing loss: Loudness and intelligibility considerations. Hear. J. 46, 31–36

  26. [34]

    Perceptual evaluation of signal-to-noise-ratio-aware dynamic range compression in hearing aids

    Kowalewski, B., Dau, T., May, T., 2020. Perceptual evaluation of signal-to-noise-ratio-aware dynamic range compression in hearing aids. Trends in Hearing 24, 2331216520930531

  27. [35]

    Evaluating the generalization of the hearing aid speech quality index (hasqi)

    Kressner, A.A., Anderson, D.V., Rozell, C.J., 2012. Evaluating the generalization of the hearing aid speech quality index (hasqi). IEEE transactions on audio, speech, and language processing 21, 407–415

  28. [36]

    Dpcrn: Dual-path convolution recurrent network for single channel speech enhancement

    Le, X., Chen, H., Chen, K., Lu, J., 2021. Dpcrn: Dual-path convolution recurrent network for single channel speech enhancement. arXiv preprint arXiv:2107.05429

  29. [37]

    Sdr–half-baked or well done?, in: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE

    Le Roux, J., Wisdom, S., Erdogan, H., Hershey, J.R., 2019. Sdr–half-baked or well done?, in: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 626–630

  30. [38]

    Hearing-loss compensation using deep neural networks: A framework and results from a listening test

    Leer, P., Jensen, J., Carney, L.H., Tan, Z.H., Østergaard, J., Bramsløw, L., 2025. Hearing-loss compensation using deep neural networks: A framework and results from a listening test. IEEE Transactions on Audio, Speech and Language Processing

  31. [39]

    Lei,T.,Hou,Z.,Hu,Y.,Yang,W.,Sun,T.,Rong,X.,Wang,D.,Chen,K.,Lu,J.,2023.Alow-latencyhybridmulti-channelspeechenhancement systemforhearingaids,in:ICASSP2023-2023IEEEInternationalConferenceonAcoustics,SpeechandSignalProcessing(ICASSP),IEEE. pp. 1–2

  32. [40]

    A dual-region speech enhancement method based on voiceprint segmentation

    Li, Y., Zhang, W.T., Lou, S.T., 2024. A dual-region speech enhancement method based on voiceprint segmentation. Neural Networks 180, 106683

  33. [41]

    Anon-invasivespeechqualityevaluationalgorithmforhearingaidswithmulti-head self-attention and audiogram-based features

    Liang,R.,Xie,Y.,Cheng,J.,Pang,C.,Schuller,B.,2024. Anon-invasivespeechqualityevaluationalgorithmforhearingaidswithmulti-head self-attention and audiogram-based features. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32, 2166–2176

  34. [42]

    1449–1457

    Lin,T.Y.,RoyChowdhury,A.,Maji,S.,2015.Bilinearcnnmodelsforfine-grainedvisualrecognition,in:ProceedingsoftheIEEEinternational conference on computer vision, pp. 1449–1457

  35. [43]

    Lin, Z., Wang, J., Li, R., Shen, F., Xuan, X., 2025. Primek-net: Multi-scale spectral learning via group prime-kernel convolutional neural networks for single channel speech enhancement, in: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Process...

  36. [44]

    Lu, Y.J., Tsao, Y., Watanabe, S., 2021. A study on speech enhancement based on diffusion probabilistic model, in: 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), IEEE. pp. 659–666

  37. [45]

    MP-SENet:Aspeechenhancementmodelwithparalleldenoisingofmagnitudeandphasespectra,in:Proc

    Lu,Y.X.,Ai,Y.,Ling,Z.H.,2023. MP-SENet:Aspeechenhancementmodelwithparalleldenoisingofmagnitudeandphasespectra,in:Proc. Interspeech, pp. 3834–3838

  38. [46]

    Luo,Y.,Chen,Z.,Yoshioka,T.,2020. Dual-pathrnn:efficientlongsequencemodelingfortime-domainsingle-channelspeechseparation,in: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 46–50

  39. [47]

    Madhu, N., Spriet, A., Jansen, S., Koning, R., Wouters, J., 2012. The potential for speech intelligibility improvement using the ideal binary mask and the ideal wiener filter in single channel noise reduction systems: Application to auditory prostheses. IEEE Transactions on Au...

  40. [48]

    Adeeplearninglossfunctionbasedontheperceptualevaluationof the speech quality

    Martin-Donas,J.M.,Gomez,A.M.,Gonzalez,J.A.,Peinado,A.M.,2018. Adeeplearninglossfunctionbasedontheperceptualevaluationof the speech quality. IEEE Signal processing letters 25, 1680–1684

  41. [49]

    Signal-to-noise-ratio-aware dynamic range compression in hearing aids

    May, T., Kowalewski, B., Dau, T., 2018. Signal-to-noise-ratio-aware dynamic range compression in hearing aids. Trends in hearing 22, 2331216518790903

  42. [50]

    A review of deep learning techniques for speech processing

    Mehrish, A., Majumder, N., Bharadwaj, R., Mihalcea, R., Poria, S., 2023. A review of deep learning techniques for speech processing. Information Fusion 99, 101869

  43. [51]

    Factorsintheeffectiveuseofhearingaidsamong subjects with age-related hearing loss: A systematic review

    Morvan,P.,Buisson-Savin,J.,Boiteux,C.,Bailly-Masson,E.,Buhl,M.,Thai-Van,H.,2024. Factorsintheeffectiveuseofhearingaidsamong subjects with age-related hearing loss: A systematic review. Journal of Clinical Medicine 13, 4027

  44. [52]

    Gated recurrent fusion to learn driving behavior from temporal multimodal data

    Narayanan, A., Siravuru, A., Dariush, B., 2020. Gated recurrent fusion to learn driving behavior from temporal multimodal data. IEEE Robotics and Automation Letters 5, 1287–1294

  45. [53]

    An integrated approach for noise reduction and dynamic range compression in hearing aids, in: 2008 16th european signal processing conference, IEEE

    Ngo, K., Doclo, S., Spriet, A., Moonen, M., Wouters, J., Jensen, S.H., 2008. An integrated approach for noise reduction and dynamic range compression in hearing aids, in: 2008 16th european signal processing conference, IEEE. pp. 1–5

  46. [54]

    Acombinedmulti-channelwienerfilter-basednoisereductionanddynamic range compression in hearing aids

    Ngo,K.,Spriet,A.,Moonen,M.,Wouters,J.,Jensen,S.H.,2012. Acombinedmulti-channelwienerfilter-basednoisereductionanddynamic range compression in hearing aids. Signal Processing 92, 417–426

  47. [55]

    Msa-dpcrn: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation, in: Proc

    Ni, Y., Pang, C., Huang, C., Zou, C., 2024. Msa-dpcrn: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation, in: Proc. Interspeech 2024, pp. 162–166

  48. [56]

    Loss functions of generative adversarial networks (gans): Opportunities and challenges

    Pan, Z., Yu, W., Wang, B., Xie, H., Sheng, V.S., Lei, J., Kwong, S., 2020. Loss functions of generative adversarial networks (gans): Opportunities and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence 4, 500–522. Ni et al.:Preprint submitted to Inf...

  49. [57]

    Librispeech: an asr corpus based on public domain audio books, in: 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), IEEE

    Panayotov, V., Chen, G., Povey, D., Khudanpur, S., 2015. Librispeech: an asr corpus based on public domain audio books, in: 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), IEEE. pp. 5206–5210

  50. [58]

    Pandey,A.,Wang,D.,2020.Denselyconnectedneuralnetworkwithdilatedconvolutionsforreal-timespeechenhancementinthetimedomain, in: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 6629–6633

  51. [59]

    The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results

    Reddy, C.K., Gopal, V., Cutler, R., Beyrami, E., Cheng, R., Dubey, H., Matusevych, S., Aichner, R., Aazami, A., Braun, S., et al., 2020. The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results. arXiv preprint arXiv:2...

  52. [60]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, 2351–2364

    Richter,J.,Welker,S.,Lemercier,J.M.,Lay,B.,Gerkmann,T.,2023.Speechenhancementanddereverberationwithdiffusion-basedgenerative models. IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, 2351–2364

  53. [61]

    Causal diffusion models for generalized speech enhancement

    Richter, J., Welker, S., Lemercier, J.M., Lay, B., Peer, T., Gerkmann, T., 2024. Causal diffusion models for generalized speech enhancement. IEEE Open Journal of Signal Processing 5, 780–789

  54. [62]

    Rix,A.W.,Beerends,J.G.,Hollier,M.P.,Hekstra,A.P.,2001. Perceptualevaluationofspeechquality(pesq)-anewmethodforspeechquality assessmentoftelephonenetworksandcodecs,in:2001IEEEinternationalconferenceonacoustics,speech,andsignalprocessing.Proceedings (Cat. No. 01CH37221), IEEE. p...

  55. [63]

    Low latency speech enhancement for hearing aids using deep filtering

    Schröter, H., Rosenkranz, T., Escalante-B, A.N., Maier, A., 2022. Low latency speech enhancement for hearing aids using deep filtering. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, 2716–2728

  56. [64]

    Measuring the acoustic effects of compression amplification on speech in noise

    Souza, P.E., Jenstad, L.M., Boike, K.T., 2006. Measuring the acoustic effects of compression amplification on speech in noise. The Journal of the Acoustical Society of America 119, 41–44

  57. [65]

    An algorithm for intelligibility prediction of time–frequency weighted noisy speech

    Taal, C.H., Hendriks, R.C., Heusdens, R., Jensen, J., 2011. An algorithm for intelligibility prediction of time–frequency weighted noisy speech. IEEE Transactions on audio, speech, and language processing 19, 2125–2136

  58. [66]

    Deep multi-frame mvdr filtering for binaural noise reduction

    Tammen, M., Doclo, S., 2022. Deep multi-frame mvdr filtering for binaural noise reduction. arxiv

  59. [67]

    A convolutional recurrent neural network for real-time speech enhancement., in: Interspeech, pp

    Tan, K., Wang, D., 2018. A convolutional recurrent neural network for real-time speech enhancement., in: Interspeech, pp. 3229–3233

  60. [68]

    Fastreal-timepersonalizedspeechenhancement:End-to-endenhancementnetwork (e3net) and knowledge distillation

    Thakker,M.,Eskimez,S.E.,Yoshioka,T.,Wang,H.,2022. Fastreal-timepersonalizedspeechenhancement:End-to-endenhancementnetwork (e3net) and knowledge distillation. arXiv preprint arXiv:2204.00771

  61. [69]

    Demand: a collection of multi-channel recordings of acoustic noise in diverse environments

    Thiemann, J., Ito, N., Vincent, E., 2013. Demand: a collection of multi-channel recordings of acoustic noise in diverse environments. (No Title)

  62. [70]

    A two-stage end-to-end system for speech-in-noise hearing aid processing

    Tu, Z., Zhang, J., Ma, N., Barker, J., et al., 2021. A two-stage end-to-end system for speech-in-noise hearing aid processing. Proc. Clarity , 3–5

  63. [71]

    Ii.noisex-92:adatabaseandanexperimenttostudytheeffectofadditivenoiseonspeechrecognitionsystems

    Varga,A.,Steeneken,H.,1993. Ii.noisex-92:adatabaseandanexperimenttostudytheeffectofadditivenoiseonspeechrecognitionsystems. Speech Commun 12, 247–251

  64. [72]

    Performancemeasurementinblindaudiosourceseparation

    Vincent,E.,Gribonval,R.,Févotte,C.,2006. Performancemeasurementinblindaudiosourceseparation. IEEEtransactionsonaudio,speech, and language processing 14, 1462–1469

  65. [73]

    Wang, Y., Zhao, Y., 2025. Heros-gan: Honed-energy regularized and optimal supervised gan for enhancing accuracy and range of low-cost accelerometers, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 21348–21356

  66. [74]

    Real-time multichannel deep speech enhancement in hearing aids: Comparing monaural and binaural processing in complex acoustic scenarios

    Westhausen, N.L., Kayser, H., Jansen, T., Meyer, B.T., 2024. Real-time multichannel deep speech enhancement in hearing aids: Comparing monaural and binaural processing in complex acoustic scenarios. IEEE/ACM Transactions on Audio, Speech, and Language Processing

  67. [75]

    Adaptive selection of local and non-local attention mechanisms for speech enhancement

    Xu, X., Tu, W., Yang, Y., 2024. Adaptive selection of local and non-local attention mechanisms for speech enhancement. Neural Networks 174, 106236

  68. [76]

    Fspen: An ultra-lightweight network for real time speech enahncment, in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE

    Yang, L., Liu, W., Meng, R., Lee, G., Baek, S., Moon, H.G., 2024. Fspen: An ultra-lightweight network for real time speech enahncment, in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 10671–10675

  69. [77]

    Anon-intrusivespeechqualityevaluationframeworkforhearingaidsbased on speech label assistance and multi-task learning strategy

    Yang,Y.,Liang,R.,Ni,Y.,Xie,Y.,Zou,C.,Schuller,B.W.,2025. Anon-intrusivespeechqualityevaluationframeworkforhearingaidsbased on speech label assistance and multi-task learning strategy. IEEE Transactions on Audio, Speech and Language Processing

  70. [78]

    Phasen: A phase-and-harmonics-aware speech enhancement network, in: Proceedings of the AAAI conference on artificial intelligence, pp

    Yin, D., Luo, C., Xiong, Z., Zeng, W., 2020. Phasen: A phase-and-harmonics-aware speech enhancement network, in: Proceedings of the AAAI conference on artificial intelligence, pp. 9458–9465

  71. [79]

    Yousefi, M., Hansen, J.H., 2021. Speaker conditioning of acoustic models using affine transformation for multi-speaker speech recognition, in: 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), IEEE. pp. 283–288

  72. [80]

    Yu, G., Li, A., Zheng, C., Guo, Y., Wang, Y., Wang, H., 2022. Dual-branch attention-in-attention transformer for single-channel speech enhancement, in: ICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP), IEEE. pp. 7847– 7851

  73. [81]

    Neural-wdrc: A deep learning wide dynamic range compression method combined with controllable noise reduction for hearing aids

    Zhang, H., Moore, B.C., Jiang, F., Diao, M., Ji, F., Li, X., Zheng, C., 2025. Neural-wdrc: A deep learning wide dynamic range compression method combined with controllable noise reduction for hearing aids. Trends in Hearing 29, 23312165241309301

  74. [82]

    Deep mcanc: A deep learning approach to multi-channel active noise control

    Zhang, H., Wang, D., 2023. Deep mcanc: A deep learning approach to multi-channel active noise control. Neural Networks 158, 318–327. Ni et al.:Preprint submitted to Information Fusion Page 24 of 27 Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and ...

  75. [2014]

    International Journal of Speech Technology 17, 53–64

    Speech enhancement with an adaptive wiener filter. International Journal of Speech Technology 17, 53–64. Ni et al.:Preprint submitted to Information Fusion Page 21 of 27 Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and Hearing Loss Compensation Ta...

Pith tools

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