Paper Archive

Automatic Chord Recognition (ACR) is constrained by the scarcity of aligned chord annotations, which are costly to acquire. At the same time, open-weight pre-trained models are more accessible than their proprietary training data. In this work, we present a two-stage training pipeline that leverages pre-trained models together with unlabeled audio. The proposed method decouples training into two stages. In the first stage, we use the pre-trained BTC model as a teacher to generate pseudo-labels for over 1,000 hours of diverse unlabeled audio and train a student model solely on these pseudo-labels. In the second stage, the student is continually trained on ground-truth labels as they become available. To prevent catastrophic forgetting of the representations learned in the first stage, we apply selective knowledge distillation (KD) from the teacher as a regularizer. In our experiments, two models (BTC, 2E1D) were used as students. In Stage 1, using only pseudo-labels, the BTC student achieves about 99% of the teacher's performance, while the 2E1D model achieves about 97% of the teacher's performance across seven standard mir_eval metrics. After continual training with labeled data in Stage 2, the resulting BTC student model consistently surpasses both the traditional supervised learning baseline and the original pre-trained teacher model across all metrics. The resulting 2E1D student model also outperforms the supervised baseline and approaches teacher-level performance, with both models demonstrating substantial gains on rare chord qualities.

Download

In this paper, we propose a new processor capable of directly changing the microdynamics of an audio signal primarily via a single dedicated user-facing parameter. The novelty of our processor is that it has built into it a measure of relative level, a short-term signal strength measurement which is robust to changes in signal macrodynamics. Consequent dynamic range processing is signal level-independent in its nature, and attempts to directly alter its observed relative level measurements. The inclusion of such a meter within our proposed processor also gives rise to a natural solution to the dynamics matching problem, where we attempt to transfer the microdynamic characteristics of one audio recording to another by means of estimating appropriate settings for the processor. We suggest a means of providing a reasonable initial guess for processor settings, followed by an efficient iterative algorithm to refine upon our estimates. Additionally, we implement the processor as a differentiable recurrent layer and show its effectiveness when wrapped around a gradient descent optimizer within a deep learning framework. Moreover, we illustrate that the proposed processor has more favorable gradient characteristics relative to a conventional dynamic range compressor. Throughout, we consider extensions of the processor, matching algorithm, and differentiable implementation for the multiband case.

Download

Virtual Analog (VA) modeling aims at digitally emulating analog audio equipment while preserving its characteristic nonlinear behavior and musical expressiveness. In the context of guitar effects, overdrive pedals represent a cornerstone of many signal chains, as they strongly contribute to the perceived dynamics, articulation, and timbral identity of the instrument. Among these, the Fulltone OCD overdrive is considered a standard in both studio and live environments, being widely adopted across rock and metal genres. In this article, we present an explicit Wave Digital (WD) model of the Fulltone OCD (v2) pedal. By exploiting the circuit topology, the MOSFETs and the germanium diode composing the asymmetric clipping stage are grouped into a single equivalent nonlinear element, enabling an explicit WD realization that avoids costly iterative solvers. The resulting nonlinear characteristic is approximated by means of a Canonical Piecewise-Linear (CPWL) function, yielding a compact and efficient explicit model suitable for real-time implementation. The proposed model is validated against reference simulations and implemented both in MATLAB and as a real-time audio plug-in using the JUCE framework.

Download

We present a self-supervised, event-based sound matching model using a differentiable extended Karplus-Strong algorithm. To avoid relying on external onset and fundamental frequency detectors, we explore training methodologies combining parameter losses on synthetic data with audio losses. We demonstrate that time-domain fractional delay interpolation provides gradient accuracy comparable to frequency-sampling while avoiding time-aliasing in highly resonant time-varying scenarios. Through systematic gradient analysis, we reveal that standard spectral losses provide no meaningful directional gradients for onset times, heavily degrading joint training. Training exclusively with parameter losses on synthetic data effectively learns fundamental frequency, timbral parameters, and onset times, but struggles to generalise to monophonic studio recordings of plucked guitar. External detectors combined with audio losses generalise best, isolating the model to timbre optimisation. While our Karplus-Strong decoder recovers interpretable parameters and naturally captures the transient characteristics of plucked guitar, Harmonics plus Noise baselines yield higher reconstruction fidelity by most metrics.

Download

This paper presents a system for morphing between audio recordings in a continuous parameter space. The proposed approach combines matrix decompositions used for audio source separation with displacement interpolation enabled by 1D optimal transport. By interpolating the spectral components obtained using nonnegative matrix factorization of the source and target signals, the system allows varying the timbre of a sound in real time, while maintaining its temporal structure. Using harmonic / percussive source separation as a pre-processing step, the system affords more detailed control of the interpolation in perceptually meaningful dimensions.

Download

Contemporary musicians commonly face the challenge of finding new, characteristic sounds that can make their compositions more distinct. They often resort to computers and algorithms, which can significantly aid in creative processes by generating unexpected material in controlled probabilistic processes. In particular, algorithms that present emergent behaviors, like genetic algorithms and cellular automata, have fostered a broad diversity of musical explorations. This article proposes an original technique for the computer-assisted creation and manipulation of sound textures. The technique uses Probabilistic Cellular Automata, which are yet seldom explored in the music domain, to blend two audio tracks into a third, different one. The proposed blending process works by dividing the source tracks into frequency bands and then associating each of the automaton’s cell to a frequency band. Only one source, chosen by the cell’s state, is active within each band. The resulting track has a non-repeating textural pattern that follows the changes in the Cellular Automata. This blending process allows the musician to choose the original material and the blend granularity, significantly changing the resulting blends. We demonstrate how to use the proposed blending process in sound design and its application in experimental and popular music.

Download

Differentiable machine learning techniques have recently proved effective for finding the parameters of Feedback Delay Networks (FDNs) so that their output matches desired perceptual qualities of target room impulse responses. However, we show that existing methods tend to fail at modeling the frequency-dependent behavior of sound energy decay that characterizes real-world environments unless properly trained. In this paper, we introduce a novel perceptual loss function based on the mel-scale energy decay relief, which generalizes the well-known time-domain energy decay curve to multiple frequency bands. We also augment the prototype FDN by incorporating differentiable wideband attenuation and output filters, and train them via backpropagation along with the other model parameters. The proposed approach improves upon existing strategies for designing and training differentiable FDNs, making it more suitable for audio processing applications where realistic and controllable artificial reverberation is desirable, such as gaming, music production, and virtual reality.

Download

Modal parameter identification for plate reverbs remains a challenging problem in virtual-analog audio effect emulation. Though neural network-based black-box approaches achieve high modeling accuracy, they generally lack interpretability and do not provide access to physically meaningful modal parameters. In this work, we present our solution to Task B of the DAFx Plate Reverb Parameter Estimation Challenge. Our method first estimates the total number of modes and then employs a Matrix Pencil (MP)-guided eigenvalue initialization strategy for a diagonal complex-valued State Space Model (SSM), which can be interpreted as a bank of parallel second-order all-pole filters. Exploiting the linearity of the resulting system, we compute the state impulse responses and replace gradient-based optimization with a closed-form least-squares estimation of the modal gains. The proposed approach enables accurate recovery of the modal parameters while maintaining an interpretable system representation.

Download

Blind source separation algorithms typically involve decorrelating time-aligned mixture signals. The usual assumption is that all sources are active at all times. However, if this is not the case, we show that the unique pattern of source activity/inactivity helps separation. Music is the most obvious example of sources exhibiting repetitive structure because it is carefully constructed. We present a novel source separation algorithm based on spatial time-time distributions that capture the repetitive structure in audio. Our method outperforms time-frequency source separation when source spectra are highly overlapping.

Download

Recent diffusion-based generative models have achieved strong results in domain-specific audio generation tasks such as speech, singing, and instrumental music synthesis. However, these models are typically specialized and do not generalize well to mixed or intermediate audio types. In this work, we adapt a diffusion-based model originally designed for multi-instrument music synthesis to voice conversion, covering both speech and singing within a unified framework. Specifically, we extend musical note-based conditioning to include phonetic posteriorgrams (PPGs) and pitch contours, and reinterpret timbre conditioning as speaker or singer identity via feature-wise linear modulation. Experiments show that the adapted model matches or surpasses a dedicated voice conversion system in terms of naturalness and performer similarity, while maintaining accurate pitch control across speech and singing. At the same time, we observe limitations in phonetic fidelity and a degradation in vocal quality when incorporating instrumental training data. Furthermore, we demonstrate that off-the-shelf feature extractors provide effective conditioning signals, enabling large-scale self-supervised training without manual annotations. These results highlight the potential of cross-domain model transfer towards unified audio generation systems capable of handling speech, singing, and music.

Download