Iterative Source Steering (ISS)

• The paper demonstrates that ISS updates demixing matrices via rank-1 steering corrections, eliminating expensive matrix inversions in IVA frameworks.
• ISS is an inverse-free update method for blind source separation that models separation as steering vector corrections, enabling efficient dynamic and selective source adaptations.
• Extensions like ISS2 and ILRMA-T-ISS leverage the ISS framework to maintain separation quality and reduce computation in real-time dereverberation and echo cancellation applications.

Searching arXiv for the specified paper and closely related work on Iterative Source Steering. Iterative Source Steering (ISS) is an inverse-free update strategy for blind source separation, most prominently in majorization-minimization-based independent vector analysis (IVA) and AuxIVA, in which the demixing or separation matrix is modified by steering-type rank-1 corrections rather than by iterative projection (IP) with explicit matrix inversion (Nakashima et al., 2022, Ikeshita et al., 2022). In common notation, the observed multichannel signal is modeled as $x=\sum_k a_k s_k = As$, the demixing matrix is $W=[w_1,\dots,w_K]^H$, and the separated signal is $y=Wx$ (Nakashima et al., 2022). Depending on the formulation, ISS is described either as updating one column of the mixing matrix $A=W^{-1}$ in each iteration or as updating the demixing matrix directly through a steering correction of the form $W_f \gets W_f - v_{kf} w_{kf}^H$ (Ikeshita et al., 2022, Nakashima et al., 2022). Its principal significance is algorithmic: it avoids matrix inversions, reduces computational cost, and in several settings exposes a steering-vector interpretation that is useful in dynamic acoustic scenes and in extensions beyond standard IVA.

1. Optimization setting and algebraic form

In MM-based IVA, ISS appears in the minimization of the surrogate objective

$$\mathcal{L}(W)=\sum_{i=1}^m w_i^{\mathsf H} V_i w_i - \log |\det W|^2,$$

where $W \in \mathbb{C}^{m\times m}$ is the separation matrix and $V_i$ are positive definite or Hermitian positive semidefinite matrices constructed in each MM iteration (Ikeshita et al., 2022). For $m \geq 3$, no algorithm has been found to obtain a global minimum of this surrogate, so practical optimization relies on block coordinate descent with closed-form updates (Ikeshita et al., 2022). Within that landscape, IP updates rows of $W$, whereas ISS updates columns of the mixing matrix $W=[w_1,\dots,w_K]^H$0 (Ikeshita et al., 2022).

In online AuxIVA and related formulations, the same mechanism is written directly as a rank-1 demixing update. The central operation is

$W=[w_1,\dots,w_K]^H$1

with source-dependent steering coefficients $W=[w_1,\dots,w_K]^H$2 chosen so that the update remains closed form and does not require matrix inversion (Nakashima et al., 2022). This makes ISS an inverse-free alternative to the standard IP update used in auxiliary-function-based IVA (Nakashima et al., 2022).

A key conceptual point established for online IVA is that this demixing-matrix update “corresponds to the update of the steering vector” (Nakashima et al., 2022). That statement is not merely interpretive terminology. It explains why ISS is useful in dynamic conditions: the update is structurally aligned with changes in the mixing geometry, rather than only with row-wise re-estimation of the demixing matrix.

2. Online AuxIVA and dynamic source motion

The 2022 online IVA formulation introduces ISS specifically for real-time blind source separation, where the demixing matrix $W=[w_1,\dots,w_K]^H$3 must be estimated frame by frame from only the current and past observations (Nakashima et al., 2022). Standard online AuxIVA inherits the per-frame inversion burden of IP, which is costly for real-time or embedded systems such as hearing aids (Nakashima et al., 2022). ISS removes that burden because it “does not require any matrix inversions,” while retaining the separation quality of AuxIVA-style processing (Nakashima et al., 2022).

The same paper emphasizes a second motivation that is specific to dynamic environments. When only part of the sources move, ISS enables the algorithm to update only the parts of the demixing model associated with the moving sources (Nakashima et al., 2022). Before convergence, the update index set $W=[w_1,\dots,w_K]^H$4 should be all sources $W=[w_1,\dots,w_K]^H$5; after convergence, once the moving sources are known, $W=[w_1,\dots,w_K]^H$6 can include only those moving-source indices (Nakashima et al., 2022). This selective updating is the mathematical expression of the steering-vector viewpoint.

The dynamic three-speaker simulation in that work used an “instantaneously move[d]” source at 30 seconds and compared IP (all), IP (one), ISS (all), and ISS (one), where “all” denotes continual updating of all sources and “one” denotes updating all sources in the first half and only the moving source in the second half (Nakashima et al., 2022). The reported runtimes and overall SDR improvements for the 60-second mixture were as follows.

Method	Runtime	SDR improvement
ISS (all)	11.001142 s	9.850487 dB
ISS (one)	8.660962 s	10.147417 dB
IP (all)	11.544285 s	8.130473 dB
IP (one)	9.004813 s	3.843095 dB

The main findings were that ISS (all) and IP (all) achieved similar separation performance with nearly overlapping segmental SDR curves, while ISS (one) maintained performance comparable to ISS (all) after the motion event and IP (one) failed to recover separation quality (Nakashima et al., 2022). The most important experimental conclusion was therefore that ISS (one) achieved almost the same SDR improvement as ISS (all) while updating only the moving source (Nakashima et al., 2022).

3. ISS, IP, and the ISS2 extension in MM-based IVA

The 2022 ISS2 paper places ISS within a broader MM and block-coordinate framework for IVA and makes the complexity trade-off explicit (Ikeshita et al., 2022). Conventional ISS, denoted ISS$W=[w_1,\dots,w_K]^H$7, updates one column of $W=[w_1,\dots,w_K]^H$8 per iteration. Conventional IP has one-row and two-row variants, IP$W=[w_1,\dots,w_K]^H$9 and IP$y=Wx$0, and IP$y=Wx$1 is described there as the current fastest IP that updates two rows of $y=Wx$2 in each iteration (Ikeshita et al., 2022).

The central limitation of conventional ISS in that comparison is not objective quality but convergence speed: although the time complexity per iteration of ISS is $y=Wx$3 times smaller than that of IP, conventional ISS converges slower than IP$y=Wx$4 (Ikeshita et al., 2022). ISS2 addresses that limitation by updating two columns of the mixing matrix $y=Wx$5 in each iteration while maintaining the small time complexity characteristic of ISS (Ikeshita et al., 2022). The paper develops a unified way of constructing ISS-type methods, denoted ISS$y=Wx$6, from which both ISS$y=Wx$7 and ISS2 are obtained systematically (Ikeshita et al., 2022).

The asymptotic per-MM-iteration complexity stated for ISS1 and ISS2 is $y=Wx$8, whereas IP1 and IP2 require $y=Wx$9 (Ikeshita et al., 2022). Numerical experiments on reverberant speech mixtures with $A=W^{-1}$0 showed that ISS2 converges in fewer MM iterations than conventional ISS and is comparable to IP$A=W^{-1}$1 (Ikeshita et al., 2022). The same experiments also noted that the lower asymptotic cost of ISS2 did not clearly translate into a wall-clock advantage in the authors’ Python implementation, which they identified as an implementation-efficiency issue rather than a contradiction of the complexity analysis (Ikeshita et al., 2022).

This places ISS in a specific methodological role. It is not merely an algebraic rewrite of IP; it is a distinct BCD family that trades inverse-based row updates for structured steering updates on $A=W^{-1}$2 or $A=W^{-1}$3, and that can itself be block-extended.

4. Unified dereverberation and separation

ISS was subsequently incorporated into joint dereverberation and blind source separation through the ILRMA-T framework (Nakashima et al., 2021). In that setting, the unified filter combines separation and dereverberation, and the objective remains the ILRMA/ILRMA-T negative log-likelihood with NMF source variance modeling (Nakashima et al., 2021). The computational problem is that ILRMA-T-IP requires two inverse matrices per source row update, which is expensive and potentially numerically fragile (Nakashima et al., 2021).

The ISS-based reformulation preserves the same probabilistic model and objective but replaces the inversion-heavy row updates with rank-1 steering updates (Nakashima et al., 2021). Two variants were proposed. ILRMA-T-ISS-JOINT alternates a separation-like ISS update for the first $A=W^{-1}$4 rows with a joint dereverberation update for the remaining taps; this reduces the inversion burden to one $A=W^{-1}$5 matrix inversion per iteration and source (Nakashima et al., 2021). ILRMA-T-ISS-SEQ updates dereverberation-related coefficients sequentially and requires no matrix inversion at all; its implementation is described as straightforward and memory efficient (Nakashima et al., 2021).

The experiments used simulated multichannel reverberant mixtures of WSJ speech in random rectangular rooms with reverberation times of 200–600 ms, CHiME3 background noise, SNR 10–30 dB, source counts 2, 3, and 4, 333 mixtures total, and a 1024-point STFT at 16 kHz with 75% overlap (Nakashima et al., 2021). The reported metrics were SI-SDR improvement, SI-SIR improvement, CD, and SRMR (Nakashima et al., 2021). Both ILRMA-T-ISS-JOINT and ILRMA-T-ISS-SEQ achieved comparable final performance to ILRMA-T-IP, dereverberation improved separation performance, and performance remained comparable even without matrix inversions (Nakashima et al., 2021). For the two-source case, the runtime of ILRMA-T-ISS-SEQ was about the same as ILRMA-T-IP, whereas for three and four sources it was much faster (Nakashima et al., 2021). The paper also reports that ISS-based methods converge slightly more slowly than ILRMA-T-IP but reach the same final performance (Nakashima et al., 2021).

A plausible implication is that ISS is especially effective when the target filter is already unified and structurally large, because inverse elimination then affects not only arithmetic cost but also numerical robustness and memory footprint.

5. Learned surrogates and accelerated descendants

ISS has also been used as the computational backbone for methods that are not limited to conventional source priors or standard AuxIVA updates. In “Surrogate Source Model Learning for Determined Source Separation,” ISS is used because of its lower complexity and lack of matrix inversion, which make it suitable for backpropagation through many iterations of AuxIVA (Scheibler et al., 2020). The paper replaces the surrogate weights $A=W^{-1}$6 in the ISS/AuxIVA update by a DNN, trains end-to-end through multiple ISS iterations, and reports large improvements in scale invariant SDR and word error rate, with the abstract highlighting as much as 36% reduction in WER (Scheibler et al., 2020). Training is done on two-speaker mixtures, and the learned approximate surrogate generalizes to mixtures of three and four speakers without modification (Scheibler et al., 2020).

A later extension inserts ISS into convolutive transfer function-based multichannel NMF. In “Accelerated Convolutive Transfer Function-Based Multichannel NMF Using Iterative Source Steering,” ISS replaces the IP demixing update in CTF-MNMF by the rank-one update $A=W^{-1}$7 (Xie et al., 30 Sep 2025). The paper states that IP has complexity $A=W^{-1}$8 and ISS has complexity $A=W^{-1}$9 (Xie et al., 30 Sep 2025). In experiments, CTF-ISS achieved comparable median SDR improvement to CTF-IP, with differences never exceeding 1 dB, and reduced runtime by about 41% for 4 microphones, 51% for 6 microphones, and 33% for 8 microphones (Xie et al., 30 Sep 2025).

An element-wise derivative of the same idea appears in semi-blind acoustic echo cancellation. “A computationally efficient semi-blind source separation based approach for nonlinear echo cancellation based on an element-wise iterative source steering” proposes EISS, which replaces a $W_f \gets W_f - v_{kf} w_{kf}^H$0 inverse in IP-based SBSS-AEC by sequential scalar steering-parameter updates (Lu et al., 2023). The complexity statement there is $W_f \gets W_f - v_{kf} w_{kf}^H$1 versus $W_f \gets W_f - v_{kf} w_{kf}^H$2 (Lu et al., 2023). On 30 signals from the AEC Challenge synthetic dataset, the reported average results were PESQ 1.89, STOI 0.93, tERLE 12.89 for IP and PESQ 1.90, STOI 0.94, tERLE 12.63 for EISS, supporting the claim that EISS achieves nearly the same AEC performance as IP while being much more computationally efficient (Lu et al., 2023).

Variant	Setting	Reported property
Learned ISS-AuxIVA surrogate	Determined source separation	Up to 36% WER reduction
CTF-MNMF-ISS	Reverberant multichannel BSS	$W_f \gets W_f - v_{kf} w_{kf}^H$3 vs. $W_f \gets W_f - v_{kf} w_{kf}^H$4
EISS	Nonlinear SBSS-AEC	$W_f \gets W_f - v_{kf} w_{kf}^H$5 vs. $W_f \gets W_f - v_{kf} w_{kf}^H$6

Across these descendants, the recurring pattern is stable: ISS is adopted when inverse-free updates are needed not only for speed, but also for differentiability, memory efficiency, or implementation on embedded hardware.

6. Terminological scope and common confusions

The acronym “ISS” is not unique to iterative source steering. In “Informed Source Separation using Iterative Reconstruction,” ISS refers to informed source separation of a single-channel mixture using Multiple Input Spectrogram Inversion, alternating time-frequency consistency enforcement and a remixing constraint (Sturmel et al., 2012). That framework begins from side information in the form of source spectrograms, initializes with mixture phase, and iteratively reconstructs waveforms; it is unrelated to the inverse-free demixing-matrix updates that define iterative source steering in IVA and related BSS models (Sturmel et al., 2012).

A second ambiguity arises from the broader use of the word “steering” in recent machine learning literature. “Self-Improving Model Steering” is about inference-time steering of LLMs and is explicitly “not exactly” iterative source steering in the narrow sense; its strongest analogue is iterative curation of prompt-response pairs through prompt ranking and contrast sampling (Zhu et al., 11 Jul 2025). “SHIFT: Steering Hidden Intermediates in Flow Transformers” is described as very close in spirit to ISS because it learns source-conditioned steering directions and applies them across diffusion timesteps, but it is not an iterative optimization method in the usual sense (Konovalova et al., 10 Apr 2026). These works are conceptually adjacent only at the level of hidden-state intervention and repeated inference-time application.

The narrow technical meaning of Iterative Source Steering in source separation therefore remains specific: an inverse-free, steering-style update rule for separation filters or demixing matrices, derived within MM or auxiliary-function frameworks and used to replace iterative projection in computationally demanding separation, dereverberation, and echo-cancellation problems (Nakashima et al., 2022, Ikeshita et al., 2022).