What Makes a Good Layer? Assessing the Layer-Wise Intrinsic Properties of Music Foundation Models

Music foundation models are commonly used as frozen audio feature extractors, yet selecting which layer to extract from remains largely heuristic. Current practice defaults to fixed depths or multi-layer fusion, with limited understanding of why certain layers transfer better across downstream tasks or how representation quality varies with depth and pre-training paradigm. We conduct a systematic layer-wise analysis of 12 music foundation models spanning three pre-training paradigms (masked modeling, autoregressive modeling, and contrastive learning), characterizing their hidden representations through intrinsic geometric and transformation-based properties. Correlating label-free representation-quality metrics with layer-wise performance across 15 downstream tasks, we find that several metrics track layer quality for genre classification, emotion recognition, automatic tagging, and beat tracking, albeit with varying strength across tasks and pre-training paradigms. However, all metrics fail on tonal tasks such as key estimation and chord recognition, indicating that no single property serves as a general proxy for representation quality across music information retrieval tasks. To address this gap, we introduce a pitch-transposition equivariance measure that captures properties missed by these standard metrics, providing a consistent indicator of tonal quality across model families. Finally, we show that intrinsic metrics can serve as effective proxies for layer selection, matching or outperforming trainable multi-layer fusion methods, particularly in limited-data settings.

Towards Robust Version Identification in the Wild: A Dataset, Benchmark, and Fine-Tuning Study

Existing datasets for musical version identification (VI) are primarily derived from curated metadata sources such as SecondHandSongs and Discogs, and are therefore dominated by professionally recorded tracks. This leads to a domain mismatch with real-world scenarios, where amateur and user-generated content is prevalent. To address this limitation, we introduce DiVers, a large-scale VI dataset comprising over 1.1 million musical versions, with train-validation-test splits compatible with established datasets such as Discogs-VI-YT, SHS100K, and Da-TACOS. In addition to standard version-level annotations, DiVers provides automatically assigned tags (e.g., instrumental, live) and segment-level predictions indicating the presence or absence of music. We evaluate the proposed dataset by training state-of-the-art VI systems. Our results show that models trained on DiVers achieve substantially improved robustness to acoustically diverse and noisy inputs, while maintaining a stable performance on cleaner, studio-quality benchmarks. We release the dataset metadata, code for its construction, and all experimental pipelines to support reproducibility.

CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation

Most Music Source Separation (MSS) models do not generalize well to live music recordings because they are trained on studio recordings alone, disregarding the venue acoustics, the speaker system's response and audience noise. We propose to bridge this gap by providing and training a model on two novel datasets. First, we present CrowdioSet: a noise dataset comprising 4800 real ambience tracks from Freesound and synthetic sing-alongs for the vocals in MUSDB18 and MOISESDB datasets, generated from zero-shot singing voice conversions. CrowdioSet enables effective audio denoising for live recordings, resulting in superior separation both in objective and subjective evaluations. Second, we introduce PaRIRset, a stereo impulse response dataset captured across 40 professional concert venues using a microphone array. Our results show that adding PaRIRset RIRs increases the performance of a MSS model compared to using real RIRs from Speech Enhancement tasks alone. We make the examples, code, model weights, PaRIRset, and CrowdioSet freely available to the public.

Evaluating Disentangled Representations for Controllable Music Generation

Recent approaches in music generation rely on disentangled representations, often labeled as structure and timbre or local and global, to enable controllable synthesis. Yet the underlying properties of these embeddings remain underexplored. In this work, we evaluate such disentangled representations in a set of music audio models for controllable generation using a probing-based framework that goes beyond standard downstream tasks. The selected models reflect diverse unsupervised disentanglement strategies, including inductive biases, data augmentations, adversarial objectives, and staged training procedures. We further isolate specific strategies to analyze their effect. Our analysis spans four key axes: informativeness, equivariance, invariance, and disentanglement, which are assessed across datasets, tasks, and controlled transformations. Our findings reveal inconsistencies between intended and actual semantics of the embeddings, suggesting that current strategies fall short of producing truly disentangled representations, and prompting a re-examination of how controllability is approached in music generation.

Benchmarking Music Autotagging with MGPHot Expert Annotations vs. Generic Tag Datasets

Music autotagging aims to automatically assign descriptive tags, such as genre, mood, or instrumentation, to audio recordings. Due to its challenges, diversity of semantic descriptions, and practical value in various applications, it has become a common downstream task for evaluating the performance of general-purpose music representations learned from audio data. We introduce a new benchmarking dataset based on the recently published MGPHot dataset, which includes expert musicological annotations, allowing for additional insights and comparisons with results obtained on common generic tag datasets. While MGPHot annotations have been shown to be useful for computational musicology, the original dataset neither includes audio nor provides evaluation setups for its use as a standardized autotagging benchmark. To address this, we provide a curated set of YouTube URLs with retrievable audio, and propose a train/val/test split for standardized evaluation, and precomputed representations for seven state-of-the-art models. Using these resources, we evaluated these models in MGPHot and standard reference tag datasets, highlighting key differences between expert and generic tag annotations. Altogether, our contributions provide a more advanced benchmarking framework for future research in music understanding.

Fractional Fourier Sound Synthesis

This paper explores the innovative application of the Fractional Fourier Transform (FrFT) in sound synthesis, highlighting its potential to redefine time-frequency analysis in audio processing. As an extension of the classical Fourier Transform, the FrFT introduces fractional order parameters, enabling a continuous interpolation between time and frequency domains and unlocking unprecedented flexibility in signal manipulation. Crucially, the FrFT also opens the possibility of directly synthesizing sounds in the alpha-domain, providing a unique framework for creating timbral and dynamic characteristics unattainable through conventional methods. This work delves into the mathematical principles of the FrFT, its historical evolution, and its capabilities for synthesizing complex audio textures. Through experimental analyses, we showcase novel sound design techniques, such as alpha-synthesis and alpha-filtering, which leverage the FrFT's time-frequency rotation properties to produce innovative sonic results. The findings affirm the FrFT's value as a transformative tool for composers, sound designers, and researchers seeking to push the boundaries of auditory creativity.

A Statistics-Driven Differentiable Approach for Sound Texture Synthesis and Analysis

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.

Automatic Estimation of Singing Voice Musical Dynamics

Musical dynamics form a core part of expressive singing voice performances. However, automatic analysis of musical dynamics for singing voice has received limited attention partly due to the scarcity of suitable datasets and a lack of clear evaluation frameworks. To address this challenge, we propose a methodology for dataset curation. Employing the proposed methodology, we compile a dataset comprising 509 musical dynamics annotated singing voice performances, aligned with 163 score files, leveraging state-of-the-art source separation and alignment techniques. The scores are sourced from the OpenScore Lieder corpus of romantic-era compositions, widely known for its wealth of expressive annotations. Utilizing the curated dataset, we train a multi-head attention based CNN model with varying window sizes to evaluate the effectiveness of estimating musical dynamics. We explored two distinct perceptually motivated input representations for the model training: log-Mel spectrum and bark-scale based features. For testing, we manually curate another dataset of 25 musical dynamics annotated performances in collaboration with a professional vocalist. We conclude through our experiments that bark-scale based features outperform log-Mel-features for the task of singing voice dynamics prediction. The dataset along with the code is shared publicly for further research on the topic.

Discogs-VI: A Musical Version Identification Dataset Based on Public Editorial Metadata

Current version identification (VI) datasets often lack sufficient size and musical diversity to train robust neural networks (NNs). Additionally, their non-representative clique size distributions prevent realistic system evaluations. To address these challenges, we explore the untapped potential of the rich editorial metadata in the Discogs music database and create a large dataset of musical versions containing about 1,900,000 versions across 348,000 cliques. Utilizing a high-precision search algorithm, we map this dataset to official music uploads on YouTube, resulting in a dataset of approximately 493,000 versions across 98,000 cliques. This dataset offers over nine times the number of cliques and over four times the number of versions than existing datasets. We demonstrate the utility of our dataset by training a baseline NN without extensive model complexities or data augmentations, which achieves competitive results on the SHS100K and Da-TACOS datasets. Our dataset, along with the tools used for its creation, the extracted audio features, and a trained model, are all publicly available online.

Heterogeneous sound classification with the Broad Sound Taxonomy and Dataset

Automatic sound classification has a wide range of applications in machine listening, enabling context-aware sound processing and understanding. This paper explores methodologies for automatically classifying heterogeneous sounds characterized by high intra-class variability. Our study evaluates the classification task using the Broad Sound Taxonomy, a two-level taxonomy comprising 28 classes designed to cover a heterogeneous range of sounds with semantic distinctions tailored for practical user applications. We construct a dataset through manual annotation to ensure accuracy, diverse representation within each class and relevance in real-world scenarios. We compare a variety of both traditional and modern machine learning approaches to establish a baseline for the task of heterogeneous sound classification. We investigate the role of input features, specifically examining how acoustically derived sound representations compare to embeddings extracted with pre-trained deep neural networks that capture both acoustic and semantic information about sounds. Experimental results illustrate that audio embeddings encoding acoustic and semantic information achieve higher accuracy in the classification task. After careful analysis of classification errors, we identify some underlying reasons for failure and propose actions to mitigate them. The paper highlights the need for deeper exploration of all stages of classification, understanding the data and adopting methodologies capable of effectively handling data complexity and generalizing in real-world sound environments.