Session 5 — Paper Presentations

Differentiable Feedback Delay Networks via Closed-Form Time-Domain Backpropagation

Antoine Salmona and Rémi Mignot

Keywords: Feedback Delay Networks, Differentiable Digital Signal Processing, Digital Audio Effects, Room Impulse Response, Deep Learning

Abstract

The Feedback Delay Network (FDN) is a widely used structure for artificial reverberation, but its Infinite Impulse Response (IIR) nature makes differentiable training non-trivial. The dominant approach, Frequency Sampling (FS), approximates gradients in the frequency domain, introducing frame-independence and possibly temporal-aliasing artifacts that can be detrimental for reverberant systems. In this paper, we derive the exact backward pass of the FDN in the time domain as a closed-form recursive formula, including a frequency-dependent feedback matrix. Unlike FS, the proposed method computes exact gradients without approximation and preserves the true IIR dynamics across successive input segments via a memory buffer, enabling streaming processing of arbitrarily long sequences. We describe a full differentiable FDN implementation and evaluate it on the task of room impulse response matching, showing that the proposed method is significantly faster than FS for comparable performance and uniquely supports streaming processing, where FS fails in this setting.

 

Multi-Scale Latent Trajectory Modelling for Audio Texture Synthesis

Aaron Dees, Susan McKeever and Seán O'Leary

Keywords: Audio texture synthesis, Multi-scale latent representations, Neural audio codec, Neural audio synthesis, Temporal dynamics modeling, Environmental audio textures

Abstract

Audio textures exhibit complex temporal dynamics that are not well represented by single-scale models. We propose a multi-scale latent trajectory approach for audio texture synthesis using latent representations extracted from a pre-trained neural audio codec. By decomposing latent trajectories into slow structural and fast stochastic components, we enable direct modeling and interpretation of the long-term temporal structure and fine-grained detail of audio textures. Our experiments on environmental audio textures suggest that the proposed method significantly improves audio quality over a single-scale baseline and achieves performance competitive with a latent diffusion model. Using perceptually motivated summary statistics, we demonstrate that generated continuations preserve texture identity. Through the proposed methods, we enable continuous texture generation from a given context and expose a latent structure that may support control over long-term dynamics.

 

From Music Audio to Expressive Dynamics Control: Transcribing MIDI CC1 from Real Music Performances via Differentiable Analysis-by-Synthesis

Yifan Xie and Mathieu Barthet

Keywords: MIDI CC1 transcription, expressive dynamics, analysis-by-synthesis

Abstract

We study automated transcription of the MIDI CC1 control trajectory from real bowed-string audio, where CC1 in sample-based orchestral libraries (e.g., Spitfire Audio's BBC SO) controls dynamics while also influencing timbral aspects through crossfading between dynamic layers. This setting lacks ground-truth CC1 annotations and suffers from a synthetic-to-real domain gap, making direct supervised learning on real recordings infeasible. We propose a differentiable analysis-by-synthesis framework that adapts a CC1 transcription network to unlabelled real audio via a learned surrogate renderer. The surrogate maps an aligned MIDI score and a CC1 trajectory to a compact expressive feature space φ consisting of root mean square (RMS) energy and spectral centroid (SC), enabling stable gradients for adaptation to real recordings without CC1 annotations. On 34 violin recordings from the URMP dataset, analysis-by-synthesis evaluation using BBC SO yields an RMS-envelope correlation of r = 0.844, improving over the model trained only on synthetic data (r = 0.675) on all 34 pieces (p < 10⁻⁶, Cohen's d = 1.90). We further observe that RMS-only adaptation can harm timbral alignment (r_CC1-SC = −0.087), while including spectral centroid in φ mitigates this degradation (Δr = +0.21, p < 10⁻⁶) without reducing RMS performance.

 

Predicting Timbre Traits for Interpretable Assessment of Musical Sound Synthesizers

Théo Chasle Cauchy, Modan Tailleur, Lindsey Reymore, Fanny Roche and Mathieu Lagrange

Keywords: Audio quality, Musical timbre modeling, Machine listening

Abstract

Quantitative assessment of neural audio synthesizers is now routinely conducted using distribution based metrics such as the Fréchet Audio Distance (FAD). Although this metric can be correlated with human perception, it offers limited interpretability beyond ranking different approaches. In this paper, we introduce a deep neural timbre trait predictor composed of a pretrained audio neural embedding CLAP, and a shallow learnable component. The latter is trained using the RWC musical instrument database and human judgments of 20 timbre descriptions (e.g., woody, percussive, rumbling, etc.) for 31 instruments. The resulting model shows strong correlation with average human ratings (r = 0.74, p < 0.001).

We then demonstrate the benefit of this predictor for evaluating the performance of TokenSynth, a neural sound synthesizer. First, the MAE computed over the set of generated sounds under different conditioning modalities of the model provides the same ranking as a FAD computed with the RWC database as a reference, suggesting that the proposed predictors are able to provide equivalent information on a distributional basis. Second, because the model is able to qualitatively analyze isolated sounds, we can determine which generated sounds could be improved and identify specific timbral dimensions that need adjustment.

 

An Evolutionary Algorithm for Automatic Guitar Amplifier Modeling

Benjamin Quiédeville, Tanguy Risset, Romain Michon and Stéphane Letz

Keywords: Evolutionary computing, Faust, Guitar amplifier modeling, Digital Signal Processing

Abstract

This paper introduces a grey-box approach for automatic guitar amplifier modeling based on evolutionary optimization. We propose a generic digital amplifier model implemented in the Faust programming language, combining domain knowledge of tube amplification with parameterized signal processing blocks. The model includes four nonlinear gain stages, tone shaping, and filtering components, resulting in a 33-parameter system. Parameter estimation is performed using an evolutionary strategy, guided by a compound loss function designed to capture spectral, temporal, and dynamic characteristics of the target amplifier. Experiments conducted on a dataset of 94 reference models demonstrate rapid convergence the proposed approach offers a computationally efficient alternative to neural methods and eliminates the need for manual parameter tuning.

 

Exploring Latent Space Audio Generation for Data Sonification

Ernst Stötzner and Andreas Nürnberger

Keywords: Sonification, Latent Space, Neural Networks, Audio Generation, Variational Autoencoders, Data Exploration

Abstract

Sonification commonly relies on parameter-mapping techniques that directly translate data values into sound synthesis parameters. In this work, we investigate an alternative approach that utilizes latent space audio generation for data sonification. We propose a method that maps data features to trajectories within the latent space of a pretrained RAVE model, enabling the generation of short audio snippets whose perceptual characteristics are controlled by the data. The approach is evaluated using the Iris dataset as an example for instances and the AirPassenger dataset as a time-series example. To assess the resulting sonifications, distances between Mel-frequency cepstral coefficient (MFCC) derived features are analyzed as a proxy for perceptual similarity. The results indicate that instances from the same class tend to produce more similar audio representations than instances from different classes, suggesting that the data structure is reflected in the generated sound. Furthermore, the time-series sonification demonstrates that temporal changes in the data are translated into perceptible variations in the audio. These findings highlight the potential of latent-space-based audio generation as a flexible approach for data-driven sound design and exploratory sonification.

 

Fifty Shades of Compression

Diemo Schwarz, Rémi Mignot, Cöme Peladeau, Haoran Sun, Étienne Démoulin, Geoffroy Peeters, Hichem Maaref and Dominique Fourer

Keywords: dynamic range compression, mixing, mastering, audio analysis, audio quality

Abstract

Audio Dynamic Range Compression (DRC) is one of the most widely used audio effects for controlling the dynamic range of an audio signal and shaping perceived audio quality. In the field of music production, particularly during mixing and mastering, efforts have been made to standardise engineering practices. However, many plugins have idiosyncratic parameters and implementations, and parameter choices remain largely empirical and depend on the individual experience and auditory perception of sound engineers. This work aims to fill this gap by introducing a proof-of-concept application that enables users to explore different sets of DRC profiles. These profiles are projected into a feature space designed to capture perceptually relevant characteristics of the output audio. By navigating this space, users can compare plugins and compression settings based on features related to perceived audio quality. We believe that this contribution can support sound engineers in selecting appropriate DRC parameters before applying compression to an audio mixture. In addition, the tool may serve pedagogical purposes by helping students and practitioners understand best practices in music production and dynamic range processing.

 

Zero-Delay Feedback for FM Cross-Modulation

Luc Döbereiner and Dario Sanfilippo

Keywords: Chaos, FM Synthesis, Feedback Systems

Abstract: FM cross-modulation --- the mutual, bidirectional phase modulation
 of two coupled oscillators --- introduces an implicit algebraic loop that the standard digital implementation resolves by inserting a one-sample delay. We show that this delay does not merely approximate the implicit system but constitutes a structurally distinct dynamical system and a persistent $f_s/4$
 spectral artefact. Drawing on Zavalishin's zero-delay feedback
 (ZDF) framework, originally developed for virtual analog filter
 design, we reformulate cross-modulation as a per-sample implicit
 equation and compare three solvers at both native sample rate and
 64x oversampling: one-sample delay, fixed-point iteration
 (FPI), and a trapezoidal Newton-Raphson method.
 FPI converges to the unique solution for beta < 1;
 oversampling eliminates the spectral artefact. The trapezoidal discretisation
 extends guaranteed per-sample convergence to beta < 2 at
 standard sample rate with eight Newton iterations. All methods
 are implemented as a SuperCollider UGen and a JUCE audio plugin;
 source code and audio examples are available in the accompanying
 repository.

 

An open source solution for live computation and broadcasting of sound descriptors

Valerio Orlandini

Keywords: sound descriptors, audiovisual interaction, sound analysis, software development, musicology, plugin, audio features

 

Abstract

The Informer is a comprehensive open-source solution for real-time extraction and distribution of audio features. The system comprises a VST3/AU/LV2 plugin, a standalone application, a VCV Rack module and externals for Max/MSP, all computing a wide range of time-domain and frequency-domain sound descriptors, with results broadcast via Open Sound Control (OSC) across all components, save for the Max/MSP externals that communicate directly with other objects. A detailed overview of the implemented descriptors, such as kurtosis, skewness, spectral centroid, flux, flatness, and more, is provided, along with their mathematical formulations and domain-specific interpretations. The accompanying C++ header-only library (with Python and JavaScript bindings) makes these computational algorithms accessible to developers, while a Max for Live device facilitates dynamic parameter modulation within Ableton Live. Built on open-source frameworks and distributed under free software licenses, The Informer bridges technical audio analysis with creative applications in music production, audiovisual performance, interactive installations, and scientific research.

 

Hierarchical Retrieval-Augmented Generation for Multimodal Music Question–Answering Using Vector-Based Semantic Retrieval

Anna-Maria Christodoulou, Kyrre Glette and Alexander Refsum Jensenius

Keywords: Retrieval-Augmented Generation (RAG), Multimodal music datasets, Cross-modal retrieval, Semantic search, Large Language Models (LLMs)

Abstract

This paper explores how hierarchical Retrieval-Augmented Generation (RAG) can support researchers' and artists' natural language interaction with complex, multimodal music datasets. Such datasets consist of heterogeneous data formats that are difficult to explore with conventional file-based or metadata-driven tools. To address this challenge, we present a lightweight prototype that combines multimodal feature extraction with vector-based semantic retrieval and a locally deployed Llama model (via Jan) to enable semantic querying and question-answering across data sources. We evaluate the system through cross-modal retrieval and dataset summarization, showing its ability to connect information across modalities and generate contextually appropriate responses. Our results show that the system reduces interaction costs by over 50\% while maintaining high retrieval accuracy. Results also suggest that the system can synthesize evidence from different data types to answer complex questions and uncover relationships within datasets that are otherwise difficult to access manually.