Artificial Audio

Example Event

Sat, 01 Jun 2030 13:00:00 +0000

Slides can be added in a few ways:

Create slides using Wowchemy’s Slides feature and link using slides parameter in the front matter of the talk file
Upload an existing slide deck to static/ and link using url_slides parameter in the front matter of the talk file
Embed your slides (e.g. Google Slides) or presentation video on this page using shortcodes.

Further event details, including page elements such as image galleries, can be added to the body of this page.

Reverberation Enhancement System

Fri, 19 Dec 2025 00:00:00 +0000

Combining digital signal processing with acoustic feedback to transform the acoustics of any space.

Concept

A reverberation enhancement system is an active system capable of controlling the room acoustics of a physical space. Microphones capture the sound present in the room, a digital signal processor enhances the signals, and loudspeakers reproduce them back into the room. This pipeline simulates changes in the geometry and absorption properties of the original space.

Your browser does not support the video tag.

The video was shot in the acoustics lab Mozart at the Fraunhofer IIS, Erlangen, Germany.
Additional information regarding the specific reverberation enhancement system adopted can be found here.

Music

Beyond engineering and room acoustics, reverberation enhancement systems unlock new possibilities for artistic expression.

Kaikuja Säiliöstä

Corresponding author: Andrea Mancianti.

A collection of site-specific immersive sound studies for brass ensemble and live electronics, written for Öljysäiliö 468, a decommissioned oil tank in East Helsinki.

Additional information regarding the musical piece and the creative process can be found here.
Paradosso

Corresponding author: Eduard Tampu.

An exploration of the influence of reverberation enhancement systems in the production and performance of electroacoustic music.

Additional information regarding the study[7] that led to the creation of this musical piece can be found here.

Tools

Open-source resources to facilitate the study and the use of reverberation enhancement systems.

Time-Varying Feedback Delay Network plugin

Real-time audio plugin implementing a Time-Varying Feedback Delay Network.

This plugin is a multi-input multi-output reverberator based on a standard feedback delay network with a time-varying feedback matrix. This architecture was originally designed and proposed for reverberation enhancement systems[2]. It is built in C++ using the JUCE framework.
DIY-RES

Setting up a reverberation enhancement system can be a challenge. DIY-RES is a guide on how to install a system using open-source software only. The proposed installation uses the Time-Varying Feedback Delay Network plugin as the system DSP.

DIY-RES offers:
- Installation instructions
  
  A written guide to setting up the transducers and using the Time-Varying Feedback Delay Network plugin in the installation.
- Signal routing templates
  
  Max/MSP and Reaper templates for routing signals from the microphones through the plugin to the loudspeakers.
DataRES

Dataset for research on reverberation enhancement systems.

Measurements from rooms with installed reverberation enhancement systems have been collected in a single open database[8]. This work facilitates the study of real-world system implementations and improves result reproducibility.
PyRES

Python library for reverberation enhancement system development and simulation.

PyRES is open-source software for testing digital signal processing architectures in reverberation enhancement systems[8]. It interfaces with DataRES, enabling simulation of real-world systems. Using FLAMO as a back-end enables the generation of custom-made DSPs as chains of elementary processing blocks. Each block operation is defined as differentiable, allowing each architecture to be trained in a DDSP fashion.
PyRES includes additional functionalities for visualization, evaluation, and auralization.

Publications

	Year	Authors	Article & accompanying material
[1]	2012	Sebastian J. Schlecht & Emanuël A. P. Habets	Reverberation enhancement from a feedback delay network perspective
[2]	2015	Sebastian J. Schlecht & Emanuël A. P. Habets	Reverberation Enhancement Systems with Time-Varying Mixing Matrices
[3]	2015	Sebastian J. Schlecht & Emanuël A. P. Habets	Time-varying feedback matrices in feedback delay networks and their application in artificial reverberation
[4]	2016	Sebastian J. Schlecht & Emanuël A. P. Habets	The stability of multichannel sound systems with time-varying mixing matrices
[5]	2024	Gian Marco De Bortoli, Karolina Prawda, & Sebastian J. Schlecht	Active Acoustics with a Phase Cancelling Modal Reverberator Accompanying material
[6]	2024	Gian Marco De Bortoli, Gloria Dal Santo, et al.	Differentiable Active Acoustics: Optimizing Stability via Gradient Descent Accompanying material
[7]	2024	Eduard Tampu	Active Acoustics: A compositional and performative approach to regenerative systems Accompanying material
[8]	2025	Gian Marco De Bortoli, Karolina Prawda, Philip Coleman, & Sebastian J. Schlecht	DataRES and PyRES: A Room Dataset and a Python Library for Reverberation Enhancement System Development, Evaluation, and Simulation Accompanying material

Acoustic Illusions for Extended Realities

Sat, 06 Sep 2025 00:00:00 +0000

Blend real and virtual sounds seamlessly by creating binaural illusions of acoustic sources.

Concept

Augmented and mixed reality systems overlay virtual sound onto the physical world. For the illusion to hold, virtual sources must be acoustically indistinguishable from real ones — a challenge that demands accurate binaural rendering, precise spatial reproduction, and an understanding of human perception.

Our research investigates when and why listeners accept virtual sounds as real, developing evaluation paradigms and rendering techniques that push the boundaries of auditory plausibility.

Transfer-Plausibility

We developed the concept of transfer-plausibility[7][13]: a rigorous framework for evaluating whether virtual sources are accepted as real when both real and virtual sounds coexist. This goes beyond traditional authenticity testing and captures the perceptual demands unique to AR/MR scenarios. Our 3AFC transfer-plausibility test proved more sensitive than alternative evaluation methods, establishing it as a standard for AR audio research.

Binaural Rendering for 6 Degrees of Freedom

Rendering spatial audio for listeners who can freely move and rotate in a space requires processing recorded Ambisonics sound fields with distance and position information[1][2]. Our work addresses source distance modeling and listener navigation through measured sound fields, enabling experiences like Inside the Quartet — an immersive experience placing the listener inside a string quartet[10].

Code: SPARTA 6DoFconv — plugin for six-degrees-of-freedom convolution with spatial room impulse responses.

Code: SRIR Interpolation Toolkit — perceptually informed interpolation of spatial room impulse responses between measurement positions.

Latency & Perceptual Thresholds

Low-latency processing is critical for maintaining the auditory illusion in real-time AR. We characterized the latency limits of head-tracked binaural rendering systems and their impact on plausibility[9], providing practical guidelines for system design.

Code: Latency Analyzer — tools for measuring binaural rendering latency.

Head-Worn Device Transparency

Wearing headphones or AR glasses disrupts the perception of real sounds. We developed methods for predicting perceptual transparency of head-worn devices[8], informing the design of passthrough processing that preserves natural listening.

Audiovisual Congruence

How do visual cues interact with spatial audio? We studied whether loudspeaker models or human avatars in VR affect localization performance[11], revealing the interplay between visual representation and spatial hearing accuracy.

Room Acoustic Memory

Can listeners remember and compare the acoustic character of spaces? Our experiments[7] investigate how accurately listeners retain room acoustic impressions, informing how quickly AR systems must adapt when transitioning between environments.

Experiences

Inside the Quartet — immersive spatial audio placing the listener inside a string quartet, demonstrating high-quality binaural rendering for musical performance[10].
Space Walk — a navigable virtual planetarium for the Oculus Quest with spatialized music[4], combining stereophonic and immersive sound spatialization.

References

	Year	Authors	Article
[1]	2018	A. Plinge, S. J. Schlecht et al.	Six-degrees-of-freedom binaural audio reproduction of first-order Ambisonics
[2]	2019	O. S. Rummukainen, S. J. Schlecht & E. A. P. Habets	Perceptual study of near-field binaural audio rendering in 6DoF VR
[3]	2020	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Fade-in control for feedback delay networks
[4]	2021	A. Mancianti, S. J. Schlecht et al.	Space Walk — visiting the solar system through an immersive sonic journey in VR
[5]	2021	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Perceptual roughness of spatially assigned sparse noise for rendering reverberation
[6]	2022	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Clearly audible room acoustical differences may not reveal where you are in a room
[7]	2022	N. Meyer-Kahlen, S. J. Schlecht et al.	Transfer-plausibility of binaural rendering with different real-world references
[8]	2022	P. Lladó, T. McKenzie, N. Meyer-Kahlen & S. J. Schlecht	Predicting perceptual transparency of head-worn devices
[9]	2023	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Latency analysis of binaural rendering systems
[10]	2023	N. Meyer-Kahlen et al.	Inside the Quartet — spatial audio experience
[11]	2024	A. Hofmann, N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Audiovisual congruence and localization in VR
[12]	2024	N. Meyer-Kahlen & S. J. Schlecht	Directional distribution of the pseudo intensity vector in anisotropic late reverberation
[13]	2024	N. Meyer-Kahlen, S. J. Schlecht et al.	Testing auditory illusions in AR: Plausibility, transfer-plausibility, and authenticity

Feedback Delay Networks for Artificial Reverberation

Sat, 06 Sep 2025 00:00:00 +0000

The fastest and most versatile way to add reverberation to sound.

Concept

Feedback delay networks (FDNs) are recursive filters that simulate the complex sound reflections in an acoustic space. A set of delay lines are connected through a feedback matrix, producing dense, natural-sounding reverberation from a compact set of parameters. Since their introduction, FDNs have become the standard building block for real-time artificial reverberators in games, music production, and spatial audio.

Our research pushes FDN theory and practice forward — from the mathematical foundations of lossless and allpass designs to practical tools for colorless, high-quality reverberation.

Lossless & Allpass Design

We established the necessary and sufficient conditions for lossless FDNs[4] and extended these results to allpass feedback delay networks[11], enabling precise spectral shaping of the reverb tail. Frequency-dependent Schroeder allpass filters[7] add further control over the spectral envelope of the reverberation.

Code: fdnToolbox — comprehensive MATLAB toolbox for FDN design and analysis, released under GNU-GPL 3.0.

Echo Density & Mixing Time

How quickly does an FDN build up a diffuse sound field? We developed the analytical characterization of echo density and mixing time in FDNs[5], a critical perceptual factor for natural-sounding reverberation. This work enables designers to predict and control the perceptual onset of diffuseness.

Scattering & Delay Feedback Matrices

Extending FDNs with scattering junctions[9] and delay-embedded feedback matrices[6] yields denser and more physically motivated reflections. These architectures bridge the gap between abstract delay networks and physical models of wave propagation.

Demo: Dense reverberation with delay feedback matrices — interactive listening examples from the delay feedback matrix framework.

Decorrelation

Maximizing output signal decorrelation[13] is essential for spatial audio and multichannel reverb rendering. Our analysis provides design guidelines for feedback matrices that produce uncorrelated output channels, directly applicable to surround and immersive audio production.

Code: fdnDecorrelation — implementation of decorrelation analysis for FDNs.

We introduced a framework for decomposing FDNs into their constituent resonant modes[8], bridging the gap between delay-network and modal descriptions of room acoustics. This decomposition enables new analysis methods and connects FDN design to the physics of room resonances. Further work on modal excitation[15] reveals how input signals interact with the FDN’s resonant structure.

Code: FDNModalShapes — visualizing and sonifying modal excitation patterns in FDNs.

Grouped FDNs

Grouped FDNs with frequency-dependent coupling[12] allow richer spectral and spatial control by connecting groups of delay lines through structured feedback. This architecture enables independent tuning of different frequency bands while maintaining the computational efficiency of FDNs.

Code: Frequency-dependent GFDN — implementation by Orchisama Das.

Colorless Reverberation

Achieving a spectrally flat (“colorless”) reverb tail is a key design goal. Using differentiable signal processing, we optimize FDN parameters via gradient descent to minimize spectral coloration[14][18]. Even tiny FDN configurations can produce high-quality colorless reverberation when properly optimized.

Code: diff-fdn-colorless — differentiable optimization of FDN coloration.

Demo: Colorless FDN listening examples — audio comparisons of optimized configurations.

Velvet Noise & Non-Exponential Decay

Dark velvet noise sequences[16] extend FDNs to model non-exponential reverberation decay, capturing the complex energy envelopes found in real rooms. The binaural dark velvet noise reverberator[16] supports spatial rendering.

Code: dark-velvet-noise-reverb — an FDN-inspired reverberator using dark velvet noise sequences, with binaural support.

Demo: Binaural DVN listening examples — spatial reverb rendering demos.

Tools

FDNTB — Feedback Delay Network Toolbox — MATLAB toolbox for FDN design and analysis. Includes feedback matrices, topologies, attenuation filters, modal decomposition, and time-varying matrices. Project page: FDNTB at DAFx 2020.
FLAMO — PyTorch library for building and optimizing differentiable audio systems. Chain differentiable gains, filters, delays, and transforms into FDN architectures and train them end-to-end. Documentation · PyPI.

References

	Year	Authors	Article
[1]	2012	S. J. Schlecht & E. A. P. Habets	Connections between parallel and serial combinations of comb filters and feedback delay networks
[2]	2015	S. J. Schlecht & E. A. P. Habets	Time-varying feedback matrices in feedback delay networks and their application in artificial reverberation
[3]	2017	S. J. Schlecht & E. A. P. Habets	Accurate reverberation time control in feedback delay networks
[4]	2017	S. J. Schlecht & E. A. P. Habets	On lossless feedback delay networks
[5]	2017	S. J. Schlecht & E. A. P. Habets	Feedback delay networks: echo density and mixing time
[6]	2019	S. J. Schlecht & E. A. P. Habets	Dense reverberation with delay feedback matrices
[7]	2019	S. J. Schlecht	Frequency-dependent Schroeder allpass filters
[8]	2019	S. J. Schlecht & E. A. P. Habets	Modal decomposition of feedback delay networks
[9]	2020	S. J. Schlecht & E. A. P. Habets	Scattering in feedback delay networks
[10]	2020	S. J. Schlecht	FDNTB: The Feedback Delay Network Toolbox
[11]	2021	S. J. Schlecht	Allpass feedback delay networks
[12]	2023	O. Das, S. J. Schlecht & E. De Sena	Grouped feedback delay networks with frequency-dependent coupling
[13]	2023	S. J. Schlecht, J. Fagerström & V. Välimäki	Decorrelation in feedback delay networks
[14]	2023	G. Dal Santo et al.	Differentiable feedback delay network for colorless reverberation
[15]	2024	S. J. Schlecht et al.	Modal excitation in feedback delay networks
[16]	2024	J. Fagerström, S. J. Schlecht & V. Välimäki	Non-exponential reverberation modeling using dark velvet noise
[17]	2024	G. Dal Santo et al.	RIR2FDN: Improved room impulse response analysis and synthesis
[18]	2025	G. Dal Santo et al.	Optimizing tiny colorless feedback delay networks
[19]	2025	G. Dal Santo et al.	FLAMO: Frequency-sampling library for audio-module optimization

Robust Measurement of Room Acoustics

Sat, 06 Sep 2025 00:00:00 +0000

Collect and clean room impulse responses in noisy and uncontrolled environments.

Concept

Measuring how a room sounds — its reverberation, reflections, and spatial characteristics — is fundamental to acoustics research, architectural design, and audio production. But real-world measurements are messy: background noise, non-stationary disturbances, and equipment limitations corrupt the signals we rely on.

Our research develops robust measurement techniques and analysis tools that extract reliable acoustic information even under difficult conditions, and makes the resulting data accessible for further processing and simulation.

Swept-Sine Measurement in Noise

Real acoustic measurements are often contaminated by non-stationary noise — footsteps, door slams, HVAC clicks. We developed the Rule of Two[5] and its short-term extension[7]: practical methods for identifying clean measurements among repeated sweeps in the presence of transient disturbances. Further work on noise removal[8] enables reliable data collection even in occupied, active spaces.

Code: mosaic_noise_removal — Python implementation of non-stationary noise removal from repeated swept-sine measurements.

Code: short-time-coherence-model — short-time coherence model for localizing noise events in sweep measurements[9].

Demo: Rule of Two listening examples — interactive demonstration of clean vs. corrupted sweep selection.

Calibrating Reverberation Models

The classic Sabine and Eyring reverberation time formulas underpin room acoustics design, yet their empirical accuracy is rarely scrutinized. We revisited these formulas using over 5,000 measurements in the variable-acoustics lab Arni[2], providing updated calibrations[4] that improve prediction accuracy for practical room design.

Data: Arni dataset — room acoustic parameter measurements across thousands of absorption configurations.

Energy Decay Analysis

Extracting reverberation parameters from room impulse responses traditionally relies on iterative curve fitting that is fragile and requires manual tuning. DecayFitNet replaces this with a lightweight neural network[3] that estimates multi-exponential energy decay parameters in a single forward pass — deterministic, fast, and validated on 20,000+ real acoustic measurements.

Common-Slope Late Reverberation Model

Real rooms rarely exhibit simple single-exponential decay. The common-slope model[6] captures multi-exponential and directional decay behavior by separating shared decay slopes from direction-dependent amplitudes. This parametric model is particularly effective for coupled rooms and complex geometries, and enables efficient real-time rendering.

Code: CommonSlopeAnalysis — MATLAB toolkit for common-slope analysis of late reverberation.

Code: blind-multi-room-model — blind estimation of multi-room acoustic models from measurements. Dataset.

Demo: Common-slope listening examples — interactive demonstrations of common-slope analysis results.

Measurement Signal Design

Designing optimal excitation signals improves measurement quality at the source. We developed two-stage filter designs[10] for measurement processing that extract cleaner impulse responses from noisy recordings.

Code: Two_stage_filter — implementation of two-stage filter design for measurement signal processing.

Demo: Two-stage filter listening examples — audio comparisons.

References

	Year	Authors	Article
[1]	2020	K. Prawda, S. J. Schlecht & V. Välimäki	Evaluation of reverberation time models with variable acoustics
[2]	2021	K. Prawda, S. J. Schlecht & V. Välimäki	Room acoustic parameters measurements in variable acoustic laboratory Arni
[3]	2022	G. Götz, S. J. Schlecht & V. Pulkki	DecayFitNet: neural network for energy decay analysis
[4]	2022	K. Prawda, S. J. Schlecht & V. Välimäki	Calibrating the Sabine and Eyring formulas
[5]	2022	K. Prawda, S. J. Schlecht & V. Välimäki	Robust selection of clean swept-sine measurements in non-stationary noise
[6]	2023	G. Götz, S. J. Schlecht & V. Pulkki	Common-slope modeling of late reverberation
[7]	2023	K. Prawda, S. J. Schlecht & V. Välimäki	Short-term Rule of Two: localizing non-stationary noise events in swept-sine measurements
[8]	2024	K. Prawda, S. J. Schlecht & V. Välimäki	Non-stationary noise removal from repeated sweep measurements
[9]	2024	K. Prawda, S. J. Schlecht & V. Välimäki	Short-time coherence model for swept-sine measurements
[10]	2024	V. Välimäki, K. Prawda & S. J. Schlecht	Two-stage filter design for measurement processing

Similarity of Sound

Sat, 06 Sep 2025 00:00:00 +0000

Computational measures of perceptual similarity of sound.

Concept

How do we quantify whether two sounds are “similar”? This question arises everywhere in audio research — from evaluating whether a synthetic reverb matches a measured room, to interpolating between spatial impulse responses, to assessing whether rendering artifacts are perceptible.

Our research develops metrics and methods for comparing sounds and acoustic environments in perceptually meaningful ways, bridging the gap between signal-level differences and what listeners actually hear.

Similarity Metrics for Late Reverberation

We developed and compared computational metrics that capture perceptual similarity between reverberant sound fields[9], going beyond simple energy-based measures to account for spectral, temporal, and spatial structure. These metrics enable objective evaluation of reverberation rendering quality.

Code: similarity-metrics-for-rirs — Python toolkit for computing and comparing similarity metrics for room impulse responses.

Demo: Reverb similarity listening examples — audio comparisons with different metrics.

Optimal Transport for Audio

Optimal transport theory provides a principled mathematical framework for comparing distributions. We apply it to quantify distances between time-frequency representations of audio signals[10] and to interpolate between spatial room impulse responses[6], yielding smooth, perceptually meaningful transitions between acoustic environments.

Code: SRIR Interpolation via Optimal Transport — perceptually informed interpolation of spatial room impulse responses using partial optimal transport.

Source Signal Similarity & Room Perception

How does the source signal itself affect our ability to hear differences between rooms? We investigated how the similarity of source material — speech, music, noise — influences listeners’ ability to distinguish between different positions in a room[8], revealing that source characteristics interact strongly with spatial perception.

Perceptual Roughness

Sparse noise signals used in efficient spatial audio rendering can introduce roughness artifacts. We quantified the perceptual roughness of spatially assigned sparse noise[2][7], establishing the thresholds where rendering artifacts become audible and guiding the design of velvet noise reverberators.

Demo: Velvet noise roughness analysis — frequency-dependent temporal roughness of velvet noise, with listening examples.

Room Acoustic Memory & Room Transitions

How accurately do listeners remember and compare room acoustics? Our experiments reveal the limits of acoustic memory[3][6], directly informing the design of systems that transition between coupled rooms. This includes work on the perceived aperture position during room transitions[4] and the perceptual analysis of directional late reverberation[1].

Data: Room transition datasets — measured acoustic data of transitions between coupled rooms.

Demo: Where you are in a room — interactive exploration of room position perception.

References

	Year	Authors	Article
[1]	2021	B. Alary, P. Massé, S. J. Schlecht et al.	Perceptual analysis of directional late reverberation
[2]	2021	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Perceptual roughness of spatially assigned sparse noise for rendering reverberation
[3]	2022	N. Meyer-Kahlen, S. J. Schlecht & T. Lokki	Clearly audible room acoustical differences may not reveal where you are in a room
[4]	2022	T. McKenzie et al.	The auditory perceived aperture position of the transition between rooms
[5]	2023	T. McKenzie et al.	Auralization of measured room transitions in VR
[6]	2023	N. Meyer-Kahlen et al.	Interpolation of spatial room impulse responses using partial optimal transport
[7]	2023	N. Meyer-Kahlen et al.	How smooth do you think I am: frequency-dependent temporal roughness of velvet noise
[8]	2023	T. McKenzie, S. J. Schlecht et al.	The role of source signal similarity in distinguishing between different positions in a room
[9]	2024	G. Dal Santo, N. Meyer-Kahlen et al.	Similarity metrics for late reverberation
[10]	2024	V. Välimäki, K. Prawda & S. J. Schlecht	Time-frequency audio similarity using optimal transport

Spatial Audio & Room Transitions

Fri, 05 Sep 2025 00:00:00 +0000

Navigate through acoustic environments with six degrees of freedom.

Concept

Real spaces are not isolated boxes — rooms connect through doorways, corridors, and open plans, and listeners move freely through them. Reproducing this experience in virtual and augmented reality requires rendering spatial audio that evolves naturally as the listener walks, turns, and transitions between coupled acoustic environments.

Our research tackles the full pipeline: from measuring and modeling the acoustics of connected spaces, to rendering smooth room transitions, to evaluating whether the result sounds convincing to listeners.

Room Transitions in Coupled Spaces

When sound travels between connected rooms, energy initially increases before decaying — creating a convex energy decay curve rather than the typical concave shape. We characterized this fade-in phenomenon[2][3] and developed auralization methods[5][9] for smooth transitions between coupled acoustic environments, including real-time rendering[11].

Code: fade-in-reverb — implementation of fade-in reverberation for multi-room environments using the common-slope model.

Demo: Fade-in reverb examples — audio demonstrations of room transitions.

Data: Room transition datasets — measured acoustic data of transitions between coupled rooms.

Common-Slope Late Reverberation Model

The common-slope model[8] decomposes late reverberation into shared decay slopes with direction-dependent amplitudes, enabling efficient, physically motivated rendering of complex reverb fields. Its extension to coupled rooms[7] and acoustic radiance transfer[12] provides a complete framework for spatial rendering in multi-room environments.

Code: CommonSlopeAnalysis — MATLAB toolkit for common-slope analysis of late reverberation.

Demo: Common-slope listening examples — interactive demonstrations of the model.

Demo: Dynamic rendering — real-time common-slope rendering for games and VR[10].

Spatial Room Impulse Response Rendering

Reproducing measured spatial room impulse responses (SRIRs) with full 6DoF listener movement requires interpolation between sparse measurement positions and anisotropic multi-slope resynthesis[4]. We developed methods for resynthesizing SRIR tails that preserve directional decay characteristics lost in conventional processing.

Code: Directional Multi-Slope RIR Resynthesis — anisotropic multi-slope decay envelope resynthesis.

Demo: Resynthesis listening examples — audio comparisons.

Code: SRIR Interpolation Toolkit — perceptually informed interpolation of SRIRs between measurement positions.

Autonomous Room Measurement

The ARTSRAM robot twin system[4a] uses autonomous Roomba-based platforms with loudspeakers and microphone arrays to collect hundreds of room impulse response measurements through random walk procedures — no prior grid planning required.

Demo: ARTSRAM project page — measurement setup and results.

Velvet Noise for Spatial Reverberation

Sparse pulse sequences (velvet noise)[6] provide an efficient basis for multichannel reverberation rendering. Dark velvet noise extends this to non-exponential decay modeling, and our perceptual studies establish quality thresholds for when rendering artifacts become audible.

Demo: Multichannel velvet noise rendering — multichannel interleaved velvet noise.

Demo: Velvet-noise FDN — spatial reverberation with velvet noise.

References

	Year	Authors	Article
[1]	2020	J. Fagerström et al.	Velvet-noise FDN for spatial reverberation
[2]	2021	T. McKenzie et al.	Acoustic analysis and dataset of transitions between coupled rooms
[3]	2021	T. McKenzie et al.	Auralization of the transition between coupled rooms
[4]	2022	C. Hold, T. McKenzie, G. Götz et al.	Resynthesis of spatial RIR tails with anisotropic multi-slope decays
[4a]	2021	G. Götz, S. J. Schlecht & V. Pulkki	ARTSRAM: Adaptive real-time spatial reverberation analysis
[5]	2023	T. McKenzie et al.	Auralization of measured room transitions in VR
[6]	2022	N. Meyer-Kahlen, S. J. Schlecht & V. Välimäki	Colours of velvet noise
[7]	2022	G. Götz et al.	Common-slope modeling of late reverberation in coupled rooms
[8]	2023	G. Götz, S. J. Schlecht & V. Pulkki	Common-slope modeling of late reverberation
[9]	2024	K. Prawda, N. Meyer-Kahlen & S. J. Schlecht	Dynamic late reverberation rendering using the common-slope model
[10]	2024	K. Y. Lee, V. Huhtala et al.	Fade-in reverberation in multi-room environments using the common-slope model
[11]	2024	P. Götz, G. Götz et al.	A common-slopes late reverberation model based on acoustic radiance transfer

Differentiable Audio Processing & Deep Learning

Thu, 04 Sep 2025 00:00:00 +0000

Bridging classical signal processing with modern machine learning for audio.

Concept

Classical audio signal processing offers transparent, interpretable algorithms — but tuning their parameters to match complex acoustic targets remains an open challenge. Deep learning brings powerful optimization, but often at the cost of interpretability and efficiency.

Our research bridges these worlds through differentiable signal processing (DDSP): embedding classical audio structures (filters, delays, feedback networks) into differentiable computation graphs that can be optimized end-to-end with gradient descent. Alongside this, we develop neural network approaches for tasks where traditional methods fall short.

Differentiable Feedback Delay Networks

Making FDN parameters differentiable allows reverberation to be optimized toward target decay, coloration, or perceptual objectives using gradient-based training. We showed that even tiny FDN configurations produce high-quality colorless reverberation when optimized this way[3][10], and developed RIR2FDN[6] for automatically synthesizing FDN configurations that match measured room impulse responses.

Code: diff-fdn-colorless — optimize FDN parameters for spectrally flat reverberation via gradient descent.

Demo: Colorless FDN examples — audio comparisons.

Code: rir2fdn — analyze measured RIRs and synthesize matching FDN configurations.

Demo: RIR2FDN project page — listening examples of RIR-to-FDN conversion.

FLAMO: Differentiable Audio Systems Library

FLAMO (Frequency-sampling Library for Audio-Module Optimization)[9] is a PyTorch library for building and optimizing differentiable linear time-invariant audio systems. It provides differentiable gains, filters (biquads, state variable filters, graphic EQs), delays, and transforms that can be chained into complex architectures and trained end-to-end.

Documentation · PyPI

Differentiable Active Acoustics

Reverberation enhancement systems form an electro-acoustic feedback loop whose stability is critical. We treat this loop as a differentiable system and optimize stability and performance via gradient descent[5], opening new possibilities for automated active acoustics design.

Demo: Differentiable active acoustics project page — demonstrations of stability optimization.

Room Impulse Response Completion

Rendering immersive audio in VR and games requires fast RIR generation. DECOR (Deep Exponential Completion Of Room impulse responses)[8] predicts late reverberation from only the early 50 ms of a measured response — an encoder-decoder network that synthesizes multi-exponential decay envelopes of filtered noise.

Demo: RIR completion project page — interactive examples.

Neural Decay Analysis

DecayFitNet[1] is a lightweight neural network that replaces brittle iterative fitting for multi-exponential energy decay estimation. Trained on synthetic data, it provides deterministic inference without manual tuning, validated on over 20,000 real acoustic measurements.

Physical Modeling with Neural Operators

Fourier neural operators[2] learn to approximate PDE solutions for physical models of musical instruments, enabling real-time sound synthesis that captures the physics of vibrating strings and resonant bodies.

Demo: FNO for physical modeling — Fourier neural operator examples.

KLANN: Knowledge-Leveraging Audio Networks

KLANN[7] integrates domain knowledge into neural network architectures for audio processing, combining the efficiency of classical signal processing structures with the flexibility of learned parameters.

Demo: KLANN examples — audio processing results.

References

	Year	Authors	Article
[1]	2022	G. Götz, S. J. Schlecht & V. Pulkki	DecayFitNet: neural network for energy decay analysis
[2]	2022	J. D. Parker, S. J. Schlecht et al.	Physical modeling with Fourier neural operators
[3]	2023	G. Dal Santo, K. Prawda et al.	Differentiable feedback delay network for colorless reverberation
[4]	2023	L. Luoma, P. Fricker & S. J. Schlecht	Deep learning for loudspeaker digital twin creation
[5]	2024	G. M. De Bortoli, G. Dal Santo et al.	Differentiable active acoustics: optimizing stability via gradient descent
[6]	2024	G. Dal Santo et al.	RIR2FDN: Improved room impulse response analysis and synthesis
[7]	2024	V. Huhtala, L. Juvela & S. J. Schlecht	KLANN: Knowledge-leveraging artificial neural network
[8]	2025	J. Lin, G. Götz & S. J. Schlecht	Deep room impulse response completion
[9]	2025	G. Dal Santo et al.	FLAMO: Frequency-sampling library for audio-module optimization
[10]	2025	G. Dal Santo, K. Prawda et al.	Optimizing tiny colorless feedback delay networks
[11]	2025	M. Scerbo, S. J. Schlecht et al.	Modeling feedback delay network output equivalences

Coding Virtual Worlds