acceptodds
Withdrawn from ICLR 2027

SFG-Mamba: Reliability-Aware Speech Frame Graph Propagation with State-Space Modeling for Speech Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-channel speech enhancement (SE) remains challenging when degradation exhibits temporal heterogeneity. Most existing methods process speech as temporally ordered sequences, without frame-level reliability estimation or direct reuse of information from reliable frames across time. To address this problem, we propose SFG-Mamba, a framework that incorporates a reliability-aware Speech Frame Graph (SFG) to model reliability-guided cross-temporal information transfer. A Speech Time-Frequency Graph Convolution (STGC) module further integrates non-local graph information with local time-frequency representations, followed by bidirectional Mamba for long-range temporal modeling. Experiments on VoiceBank+DEMAND and DNS demonstrate the effectiveness of SFG-Mamba, achieving PESQ scores of 3.57 and 3.52, respectively. Furthermore, additional experiments are designed to investigate the performance drop under increasing temporal heterogeneity. SFG-Mamba achieves the best performance among the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
speech enhancement, graph neural networks, state space models, cross-frame information reuse, non-uniform speech degradation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ML6GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SFG-Mamba, a single-channel speech enhancement framework aimed at non-uniform temporal degradation. The core mechanism is a reliability-aware Speech Frame Graph (SFG): per-frame reliability is estimated from the noisy signal via MCRA-based noise estimation (Eq. 5–6), the top-K reliable frames are selected as graph sources and frames below a threshold τ_q as targets (Eq. 7–8), and information propagates along directed edges weighted by cosine similarity and a learnable pairwise gate (Eq. 9–11). The SFG branch is fused with a local convolutional branch inside the STGC module (Eq. 13), and the result is processed by a bidirectional Mamba (Eq. 14–15) before magnitude/phase decoding. The paper also introduces Temporal Degradation Dispersion (TDD) as a proxy for temporal heterogeneity and runs a controlled evaluation at fixed global SNR (0 dB) where only the temporal arrangement of noise varies (Section 4.3). Experiments on VoiceBank+DEMAND (PESQ 3.57) and DNS (PESQ 3.52) show gains over listed baselines, supported by ablations (Table 2), a K-sensitivity sweep (Figure 5), and spectrogram visualizations (Figure 3).

Critical strengths
  1. The TDD-controlled evaluation (Section 4.3, Figure 4) is the strongest part of the paper. Constructing noise as K=10 burst events plus a continuous background component, holding total noise energy fixed at 0 dB global SNR, and sweeping only the temporal clustering is a clean way to isolate heterogeneity from noise level. Most SE evaluations conflate the two by varying global SNR, so this setup actually answers a question the field leaves open. If the confound I raise below were fixed, this section alone would carry real weight.

  2. The reliability estimation is reference-free and deployment-practical. Using MCRA on the noisy power spectrum (Eq. 5–6) to score frames means the graph construction needs no clean speech at inference, which is the right setting for a real enhancement system. The directed bipartite design (sources = reliable, targets = degraded) is also a sensible inductive bias — it prevents noise-dominated frames from acting as propagation sources, which the related graph SE methods (Tzirakis 2021, Chau 2024) do not explicitly prevent.

Critical weaknesses
  1. The TDD experiment has a built-in confound that undermines the paper's central claim. Section 4.3 ties the background energy fraction to TDD via f(TDD) = 0.10·TDD⁴, so as TDD increases the noise becomes not only more temporally clustered but also more dominated by continuous background energy relative to bursts. The performance drops in Figure 4 could therefore be driven by the changing burst/background mix rather than by temporal heterogeneity itself. The paper calls this coupling "intentional" and argues it mimics real recordings, but that is an assertion, not evidence — no experiment disentangles the two factors, and without one the claim that "non-uniform degradation, quantified by TDD, significantly affects speech enhancement performance" (Section 4.3) is not fully supported. This is fixable (e.g., a 2×2 sweep over clustering and background fraction), but as written it weakens the most interesting part of the paper.

  2. The novelty is incremental and the paper doesn't clearly articulate what is fundamentally new. Mamba for speech enhancement already exists (Audio Mamba, Erol 2024; aTENNuate, Pei 2025; Dual-Path Mamba, Jiang 2025), and graph-based SE exists (Tzirakis 2021, Chau 2024, Wang 2025, Saleem 2025). The SFG's only distinguishing ingredient is reliability-based source selection on top of a standard cosine-similarity graph — a fairly small delta over the vanilla GCN baseline in Table 2 (3.10 vs. 3.57 PESQ, though that gap conflates the graph design with the directed/reliability structure). The related work (Section 2) concedes that prior methods already "learn or construct acoustic topology"; the paper needs a sharper statement of what reliability-aware directed propagation buys beyond that, ideally with an experiment that isolates it from the graph structure itself.

Questions for authors
  1. Can you decouple the burst/background energy ratio from TDD in the Section 4.3 setup? As written, f(TDD) = 0.10·TDD⁴ entangles them — a 2×2 sweep (clustering × background fraction) would make the robustness claim much stronger.
  2. What value of τ_q (Eq. 7) was used, and how sensitive are the results to it? K is swept in Figure 5 but τ_q is never reported.
  3. What are the parameter count, FLOPs, and per-utterance runtime of SFG-Mamba compared to MPSENet and HRLF-Net? No complexity analysis is provided.
  4. The MCRA noise estimate is non-differentiable — is it precomputed per utterance and treated as a constant during training? If so, how is it batched, and does the model see the same estimate at train and test time?
  5. Figure 1(b) plots performance against TDD — is this on synthetic data only? Is there any measurement of TDD on real recordings to back the claim that real-world degradation is highly non-uniform?
  6. The ablation replaces SFG with a vanilla GCN (Table 2), but a vanilla GCN is not the same as "no graph." Have you tried a no-graph variant (LTF + Bi-Mamba only) to separate the contribution of having a graph at all from the contribution of the reliability-aware design?
Reviewer_QJ2yOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SFG-Mamba combines reliability-guided speech frame graph propagation, local time-frequency processing, and bidirectional Mamba for single-channel speech enhancement. It reports improvements on VoiceBank+DEMAND and DNS and investigates robustness to temporally non-uniform degradation.

Critical strengths

The motivation of using reliable frames to support degraded regions is reasonable. The graph provides an potentially interpretable information-transfer mechanism, and the reported benchmark results are encouraging.

Critical weaknesses
  1. The temporal-heterogeneity experiment changes both burst placement and background-noise energy, preventing isolation of the effect of temporal irregularity. 2 The reliability estimates lack direct validation under transient interference, and the ablations do not isolate the benefit of reliability-based source selection.
  2. The ablations and temporal-heterogeneity analysis are limited to VoiceBank+DEMAND, while the DNS protocol is insufficiently specified to assess reproducibility and generalizability.
  3. The comparisons omit the extended MP-SENet results, weakening the broad performance claims unless differences in evaluation protocols are clarified.
  4. The claim that sequence-based models cannot directly exploit distant frames is overstated, since attention already supports such interactions.
  5. The description does not clearly specify backward-output alignment in Bi-Mamba or source-node updates across graph layers.
Questions for authors
  1. Can the temporal-heterogeneity experiment vary only burst positions while keeping the background signal and burst characteristics fixed?
  2. How accurate are the reliability estimates during transient interference, and how does reliability-based selection compare with random or similarity-only selection?
  3. Which baseline results were reproduced under matched conditions, and what exact DNS training and test protocols were used?
  4. Is the backward Mamba output restored to the original temporal order before fusion, and how are source features maintained across graph layers?
Reviewer_C1dcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a single-channel SE model that aims to improve on temporally non-uniform noise degradation. It is based on MPSENet, using the same pre-processing, magnitude/phase decoders, and losses. It replaces the DenseNet encoder with a graph-based encoder (STGC) and the Conformer with BiMamba.

The encoder constructs a bipartite graph from frames with high estimated reliability (sources) to frames with low estimated reliability (targets). Edges are weighted by source reliability, cosine similarity between source and target frame, and a learned pairwise gate. Each target frame is updated as teh weighted sum of projected source frames.

Reliability is estimated per TF-bin using MCRA noise power estimation. From the noise power, the a priori SNR is estimated and used to compute a frame-level score. The paper also introduces Temporal Degradation Dispersion (TDD) as a proxy metric to quantify temporal non-uniformity of degradation. It is the coefficient of variation of time intervals between noise events.

Critical strengths
  • The authors address the important problem of "non-uniform degradation" in speech enhancement. The proposed Temporal Degradation Dispersion (TDD) is a reasonable proxy metric for such noise, and figure 1 (b) shows that speech enhancement metrics are sensitive to increasing TDD.
  • The proposed SFG introduces an interesting idea of propagating information from relatively clean (hight-reliability) frames to heavily degraded frames. The graph structure is a good choice to connect frames based on acoustic similarity, and, potentially, can improve the reconstruction of corrupted speech segments using information from distant temporal regions.
Critical weaknesses
  • Ablations are missing to show the actual benefit of the proposed reliability-aware frame selection.

  • The baselines in the main results table (table 1) are outdated and not sufficiently relevant. There are at least nine (!) Mamba-based speech-enhancement models, published in 2024 or later that outperforms the proposed SFG-Mamba method. Even the very first SEMamba papers was not mentioned in the current paper. One might assume that the authors selected baselines based on other criteria, such as graph-based speech enhancement methods or methods specifically designed for impulsive noise. However, none of the selected methods uses or investigates these approaches. Some examples for mamba papers with higher PESQ performance:

  • The authors do not report +- standard deviations or confidence intervals for the results. Without these, it is difficult to assess whether the reported improvements are statistically meaningful (for example, given the subtle changes, the "benefits" in Figure 5 may just be within the standard deviation, or in other words: noise).

  • The importance of the new method is not convincingly demonstrated. A more appropriate ablation for Figure 5 would be to use K randomly selected frames instead of the Top-K frames selected by the proposed method.

  • For table 2, the authors should isolate the SFG and LTF modules by removing each of them separately as well as both together.

  • The description of the main method is hard to follow. For instance, what was used for the training set: DNS or VoiceBank-Demand datasets? A clearer explanation on loss functions (Do authors use the PESQ-loss?)?

  • Not all hyperparameters are specified. Check for instance: power-law compression cc. Number of mamba blocks. Reliability Threshold. Hidden size in SFG layers, threshold τq\tau_q

  • L74 - L77: Unsupported claim that processing speech as temporal sequence weakens contribution of highly informative frame. Transformers use attention for non-local information transfer, Mamba uses time-step modulation to suppress integration of irrelevant input features.

  • Please check if Figure 2 matches the description in sections 3.1.2 and 3.2

  • L352: Its unclear why different metrics are used for the datasets, and why there are no non-intrusive metrics for DNS

Questions for authors

See my concerns above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.