Sebastian J. Schlecht

Room impulse response (RIR) estimation naturally arises as a class of inverse problems, including denoising and deconvolution. While recent approaches often rely on supervised learning or learned generative priors, such methods require large amounts of training data and may generalize poorly outside the training distribution. In this work, we present RIRFlow, a training-free Bayesian framework for RIR inverse problems using flow matching. We derive a flow-consistent analytic prior from the statistical structure of RIRs, eliminating the need for data-driven priors. Specifically, we model RIR as a Gaussian process with exponentially decaying variance, which yields a closed-form Wiener denoiser. This analytic denoiser is integrated as a prior in an existing flow-based inverse solver, where inverse problems are solved via guided posterior sampling. Furthermore, we extend the solver to nonlinear and non-Gaussian inverse problems via a local Gaussian approximation of the guided posterior, and empirically demonstrate that this approximation remains effective in practice. Experiments on real RIRs across different inverse problems demonstrate robust performance, highlighting the effectiveness of combining a classic RIR model with the recent flow-based generative inference.

Active acoustic enhancement systems (AAESs) use microphones, loudspeakers and electronic processing to modify the reverberation of a space, offering flexible and cost-effective alternatives to passive variable acoustics. These systems can extend the reverberation time of a space and modify perceived characteristics such as wall distance, diffuseness and intimacy. In this article, the current literature is discussed, and common conditions of AAESs are demonstrated using simulations to help researchers to establish a comprehensive understanding of the field. A general model is first defined to approximate any AAES as a linear, time-invariant system of transfer functions. This is used to analyse the general stability condition, which is valuable for system tuning and prediction. The three main topologies of AAESs are presented, namely, in-line, regenerative and hybrid systems, describing the fundamental differences as well as the nuances of commercial implementations with a focus on signal processing techniques. Articles investigating AAESs have been summarised to allow readers to gauge the coverage of experimental research to date. The simulated contribution serves as an exploratory environment to compare AAES conditions, where code and audio examples are available online. Promising future trajectories are identified involving machine learning, artefact perception and expressive performance.

Rendering dynamic reverberation in a complicated acoustic space for moving sources and listeners is challenging but crucial for enhancing user immersion in extended-reality (XR) applications. Capturing spatially varying room impulse responses (RIRs) is costly and often impractical. Moreover, dynamic convolution with measured RIRs is computationally expensive with high memory demands, typically not available on wearable computing devices. Grouped Feedback Delay Networks (GFDNs), on the other hand, allow efficient rendering of coupled room acoustics. However, its parameters need to be tuned to match the reverberation profile of a coupled space. In this work, we propose the concept of Differentiable GFDNs (DiffGFDNs), which have tunable parameters that are optimised to match the late reverberation profile of a set of RIRs captured from a space that exhibits multi-slope decay. Once trained on a finite set of measurements, the DiffGFDN interpolates to unmeasured locations in the space. We propose a parallel processing pipeline that has multiple DiffGFDNs with frequency-independent parameters processing each octave band. The parameters of the DiffGFDN can be updated rapidly during inferencing as sources and listeners move. We evaluate the proposed architecture against the Common Slopes (CS) model on a dataset of RIRs for three coupled rooms. The proposed architecture generates multi-slope late reverberation with low memory and computational requirements, achieving a better energy decay relief (EDR) error and slightly worse octave-band energy decay curve (EDC) errors compared to the CS model. Furthermore, DiffGFDN requires an order of magnitude fewer floating-point operations per sample than the CS renderer.

This paper addresses the analytic Procrustes problem, which aims to find the best least-squares paraunitary approximation of a square matrix of analytic transfer functions, or the best paraunitary transformation between two rectangular analytic matrices. This is accomplished by generalising the Procrustes solution from ordinary matrices to the case of matrices of analytic functions via their analytic singular value decomposition (SVD). Different from the ordinary matrix case, the analytic SVD is not restricted to singular values being nonnegative. In the case that singular values do not possess any zero crossings, we can find an analytic paraunitary matrix analogously to the standard Procrustes approach. In the case that singular values exhibit any zero crossings, the solution does not only depend on the left- and right-singular vectors, but also on a discontinuous and hence non-analytic switching function that forces those analytic singular values to become nonnegative real. We show that a close approximation of this switching function can be achieved via a complex-valued allpass filter, for which we suggest a new suitable design to minimise the overall least squares error of the fit. In addition, we propose a DFT domain algorithm to approximate this polynomial Procrustes solution, which avoids ambiguities in the analytic SVD, and possesses proven convergence. Generally, this solution requires a delay for causality, and this delay grows with the approximation order. Examples and simulations demonstrate our proposed method.

Spatial audio systems are typically evaluated in comparative listening tests using the same source signal for each condition {such as ABX: ITU-R BS.1116-3 [(2015a) Methods for the Subjective Assessment of Small Impairments in Audio Systems (International Telecommunication Union, Geneva, Switzerland)] and multiple stimulus with hidden reference and anchor ITU-R BS.1534-3 [(2015b) Methods for the Subjective Assessment of Intermediate Quality Level of Audio Systems (International Telecommunication Union, Geneva, Switzerland)]}. However, in augmented reality (AR) scenarios, it is infeasible that the same sound source would exist at the same position in space, both real and virtual; instead, each sound source will emit a different signal. To investigate this discrepancy, a perceptual study is conducted on the effect of source signal similarity when distinguishing different room acoustics conditions. Specifically, these conditions are binaural room impulse responses measured at different distances from the source, modified to all use the same direct sound. Three classes of source signal are investigated in a three-alternative forced choice paradigm: the same speech signal for all conditions, the same speaker but a different sentence for each condition, and a different speaker and a different sentence for each condition. Results show that using different speech recordings significantly reduces the ability to identify differences in room acoustics. This suggests that spatial audio system fidelity requirements could vary depending on the source signals used in the target application; AR audio evaluation should use different signals for comparisons.

In complex acoustic environments such as multiple connected rooms, reverberation is highly dependent on the posi tions of sound sources and listeners — not only in terms of early reflections, but late reverberation as well. Modeling this positional dependency accurately is important for immersive, interactive applications such as virtual reality, augmented reality, and video games, where reverberation needs to be adapted in real time as sound sources and listeners move. The recently proposed modal decomposition of acoustic radiance transfer (MoD-ART) method can evaluate position-dependent late reverberation characteristics in real time, based on physical properties of the modeled environment, and it is specifically designed for complex acoustic environments. The reverberation characteristics' auralization (i.e. their application to audio signals) can be accomplished either with convolution or with delay-based reverberators. In this paper, we propose a method to auralize late reverberation efficiently in the presence of multiple sound sources and listeners, based on the MoD-ART model. The proposed method inherits the favorable complexity scaling of MoD-ART's modeling and extends it to the aspect of auralization, enabling the rendering of late reverberation in scenarios with hundreds of interactive sound sources and listeners. Furthermore, the proposed method can model fully dynamic scenarios (meaning both sources and listeners may move) correctly and with no rendering latency.

The presence of unavoidable background noise limits the signal-to-noise ratio in measured room impulse responses (RIRs). A common solution is to crop the RIR to the time interval where the signal dominates the background noise, but finding the correct onset and truncation points is challenging. It usually requires estimating the sound decay rate and noise floor, which is burdened with uncertainty. In this study, we propose an RIR cropping method based on the covariance between two repeated RIRs and its inherent monotonicity. Evaluation on measured RIRs shows the proposed method is highly robust in different scenarios and outperforms state-of-the-art algorithms.

Modeling late reverberation in real-time interactive applications is a challenging task when multiple sound sources and listeners are present in the same environment. This is especially problematic when the environment is geometrically complex and/or features uneven energy absorption (e.g. coupled volumes), because in such cases the late reverberation is dependent on the sound sources' and listeners' positions, and therefore must be adapted to their movements in real time.We present a novel approach to the task, named modal decomposition of acoustic radiance transfer (MoD-ART), which can handle highly complex scenarios with efficiency. The approach is based on the geometrical acousticsmethod of acoustic radiance transfer, fromwhich we extract a set of energy decaymodes and their positional relationships with sources and listeners. In this paper, we describe the physical and mathematical significance of MoD-ART, highlighting its advantages and applicability to different scenarios. Through an analysis of the method's computational complexity, we show that it compares very favorably with ray-tracing.We also present simulation results showing thatMoD-ART can capture multiple decay slopes and flutter echoes.

Experiments testing sound for augmented reality can involve real and virtual sound sources. Paradigms are either based on rating various acoustic attributes or testing whether a virtual sound source is believed to be real (i.e., evokes an auditory illusion). This study compares four experimental designs indicating such illusions. The first is an ABX task suitable for evaluation under the authenticity paradigm. The second is a Yes/No task, as proposed to evaluate plausibility. The third is a three-alternative-forced-choice (3AFC) task using different source signals for real and virtual, proposed to evaluate transfer-plausibility. Finally, a 2AFC task was tested. The renderings compared in the tests encompassed mismatches between real and virtual room acoustics. Results confirm that authenticity is hard to achieve under nonideal conditions, and ceiling effects occur because differences are always detected. Thus, the other paradigms are better suited for evaluating practical augmented reality audio systems. Detection analysis further shows that the 3AFC transfer-plausibility test is more sensitive than the 2AFC task. Moreover, participants are more sensitive to differences between real and virtual sources in the Yes/No task than theory predicts. This contribution aims to aid in selecting experimental paradigms in future experiments regarding perceptual and technical requirements for sound in augmented reality.