Paper Archive

Machine learning models have become ubiquitous in modeling analog audio devices. Expanding on this line of research, our study focuses on Voltage-Controlled Oscillators of analog synthesizers. We employ black box autoregressive artificial neural networks to model the typical analog waveshapes, including triangle, square, and sawtooth. The models can be conditioned on wave frequency and type, enabling the generation of pitch envelopes and morphing across waveshapes. We conduct evaluations on both synthetic and analog datasets to assess the accuracy of various architectural variants. The LSTM variant performed better, although lower frequency ranges present particular challenges.

Download

Machine learning for sound generation is rapidly expanding within the computer music community. However, most datasets used to train models are built from field recordings, foley sounds, instrumental notes, or commercial music. This presents a significant limitation for composers working in acousmatic and electroacoustic music, who require datasets tailored to their creative processes. To address this gap, we introduce the SCHAEFFER Dataset (Spectromorphological Corpus of Human-annotated Audio with Electroacoustic Features For Experimental Research), a curated collection of 1000 sound objects designed and annotated by composers and students of electroacoustic composition. The dataset, distributed under Creative Commons licenses, features annotations combining technical and poetic descriptions, alongside classifications based on pre-defined spectromorphological categories.

Download

Modal synthesis methods are a long-standing approach for modelling distributed musical systems. In some cases extensions are possible in order to handle geometric nonlinearities. One such case is the high-amplitude vibration of a string, where geometric nonlinear effects lead to perceptually important effects including pitch glides and a dependence of brightness on striking amplitude. A modal decomposition leads to a coupled nonlinear system of ordinary differential equations. Recent work in applied machine learning approaches (in particular neural ordinary differential equations) has been used to model lumped dynamic systems such as electronic circuits automatically from data. In this work, we examine how modal decomposition can be combined with neural ordinary differential equations for modelling distributed musical systems. The proposed model leverages the analytical solution for linear vibration of system’s modes and employs a neural network to account for nonlinear dynamic behaviour. Physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the network architecture. As an initial proof of concept, we generate synthetic data for a nonlinear transverse string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented.

Download

This paper investigates the synthesis of audio textures and the structure of generative latent spaces using Variational Autoencoders (VAEs) within two paradigms of neural audio synthesis: DSP-inspired and data-driven approaches. For each paradigm, we propose VAE-based frameworks that allow fine-grained temporal control. We introduce datasets across three categories of environmental sounds to support our investigations. We evaluate and compare the models’ reconstruction performance using objective metrics, and investigate their generative capabilities and latent space structure through latent space interpolations.

Download

This work focuses on developing an artistic tool that performs an unsupervised mapping between text and sound, converting an input text string into a series of sounds from a given sound corpus. With the use of a pre-trained sound embedding model and a separate, pre-trained text embedding model, the goal is to find a mapping between the two feature spaces. Our approach is unsupervised which allows any sound corpus to be used with the system. The tool performs the task of text-to-sound retrieval, creating a soundfile in which each word in the text input is mapped to a single sound in the corpus, and the resulting sounds are concatenated to play sequentially. We experiment with three different mapping methods, and perform quantitative and qualitative evaluations on the outputs. Our results demonstrate the potential of unsupervised methods for creative applications in text-to-sound mapping.

Download

Conventional Deep Learning (DL) approaches to Infinite Impulse Response (IIR) filter coefficients estimation from arbitrary frequency response are quite limited. They often suffer from inefficiencies such as tight training requirements, high complexity, and limited accuracy. As an alternative, in this paper, we explore the use of Kolmogorov-Arnold Networks (KANs) to predict the IIR filter—specifically biquad coefficients—effectively. By leveraging the high interpretability and accuracy of KANs, we achieve smooth coefficients’ optimization. Furthermore, by constraining the search space and exploring different loss functions, we demonstrate improved performance in speed and accuracy. Our approach is evaluated against other existing differentiable IIR filter solutions. The results show significant advantages of KANs over existing methods, offering steadier convergences and more accurate results. This offers new possibilities for integrating digital infinite impulse response (IIR) filters into deep-learning frameworks.

Download

In this paper, we present an approach to the neural modeling of overdrive guitar pedals with conditioning from a cross-circuit and cross-setting latent space. The resulting network models the behavior of multiple overdrive pedals across different settings, offering continuous morphing between real configurations and hybrid behaviors. Compact conditioning spaces are obtained through unsupervised training of a variational autoencoder with adversarial training, resulting in accurate reconstruction performance across different sets of pedals. We then compare three Hyper-Recurrent architectures for processing, including dynamic and static HyperRNNs, and a smaller model for real-time processing. Additionally, we present pOD-set, a new open dataset including recordings of 27 analog overdrive pedals, each with 36 gain and tone parameter combinations totaling over 97 hours of recordings. Precise parameter setting was achieved through a custom-deployed recording robot.

Download

Deep State Space Models (SSMs) have shown remarkable performance in long-sequence reasoning tasks, such as raw audio classification, and audio generation. This paper introduces PianoSSM, an end-to-end deep SSM neural network architecture designed to synthesize raw piano audio directly from MIDI input. The network requires no intermediate representations or domainspecific expert knowledge, simplifying training and improving accessibility. Quantitative evaluations on the MAESTRO dataset show that Piano-SSM achieves a Multi-Scale Spectral Loss (MSSL) of 7.02 at 16kHz, outperforming DDSP-Piano v1 with a MSSL of 7.09. At 24kHz, Piano-SSM maintains competitive performance with an MSSL of 6.75, closely matching DDSP-Piano v2’s result of 6.58. Evaluations on the MAPS dataset achieve an MSSL score of 8.23, which demonstrates the generalization capability even when training with very limited data. Further analysis highlights Piano-SSM’s ability to train on high sampling-rate audio while synthesizing audio at lower sampling rates, explicitly linking performance loss to aliasing effects. Additionally, the proposed model facilitates real-time causal inference through a custom C++17 header-only implementation. Using an Intel Core i712700 processor at 4.5GHz, with single core inference, allows synthesizing one second of audio at 44.1kHz in 0.44s with a workload of 23.1GFLOPS/s and an 10.1µs input/output delay with the largest network. While the smallest network at 16kHz only needs 0.04s with 2.3GFLOP/s and 2.6µs input/output delay. These results underscore Piano-SSM’s practical utility and efficiency in real-time audio synthesis applications.

Download

In this work, we introduce TexStat, a novel loss function specifically designed for the analysis and synthesis of texture sounds characterized by stochastic structure and perceptual stationarity. Drawing inspiration from the statistical and perceptual framework of McDermott and Simoncelli, TexStat identifies similarities between signals belonging to the same texture category without relying on temporal structure. We also propose using TexStat as a validation metric alongside Frechet Audio Distances (FAD) to evaluate texture sound synthesis models. In addition to TexStat, we present TexEnv, an efficient, lightweight and differentiable texture sound synthesizer that generates audio by imposing amplitude envelopes on filtered noise. We further integrate these components into TexDSP, a DDSP-inspired generative model tailored for texture sounds. Through extensive experiments across various texture sound types, we demonstrate that TexStat is perceptually meaningful, time-invariant, and robust to noise, features that make it effective both as a loss function for generative tasks and as a validation metric. All tools and code are provided as open-source contributions and our PyTorch implementations are efficient, differentiable, and highly configurable, enabling its use in both generative tasks and as a perceptually grounded evaluation metric.

Download

Neural networks have been applied within the Wave Digital Filter (WDF) framework as data-driven models for nonlinear multi-port circuit elements. Conventionally, these models are trained on wave variables obtained by sampling the current-voltage characteristic of the considered nonlinear element before being incorporated into the circuit WDF implementation. However, isolating multi-port elements for this process can be challenging, as their nonlinear behavior often depends on dynamic effects that emerge from interactions with the surrounding circuit. In this paper, we propose a novel approach for training neural models of nonlinear multi-port elements directly within a circuit’s Wave Digital (WD) discretetime implementation, relying solely on circuit input-output voltage measurements. Exploiting the differentiability of WD simulations, we embed the neural network into the simulation process and optimize its parameters using gradient-based methods by minimizing a loss function defined over the circuit output voltage. Experimental results demonstrate the effectiveness of the proposed approach in accurately capturing the nonlinear circuit behavior, while preserving the interpretability and modularity of WDFs.

Download