Philippe Esling

Recently, research on audio foundation models has witnessed notable advances, as illustrated by the ever improving results on complex downstream tasks. Subsequently, those pretrained networks have quickly been used for various audio applications. These improvements have however resulted in a considerable increase both in size and complexity of these models. Along the environmental concerns this issue raises, this prevents the deployment of such networks on consumer-level devices, and precludes their use for real-time applications. Moreover, this appears contradictory with the specificity of the tasks for which these models are used, which are often simpler compared to extracting a rich, multi-purpose representation from any type of audio data. In this paper, we address this issue with a simple, yet effective method to extract lightweight specialist subnetworks from large foundation models. Specifically, we introduce learnable binary masks in-between the layers of a pretrained representation model. When training the end-to-end model on a downstream task, we add a sparsity-inducing loss to the overall objective, hence learning a compact subnetwork specialized on a single task. Importantly, the weights of the foundation model are kept frozen, resulting into low additional training costs. Once trained, the masked computational units can then be removed from the network, implying significant performance gains. We assess our method on three widespread audio foundation models, each based on a different backbone architecture, and illustrate its effectiveness on common audio representation evaluation tasks, as well as its versatility on both speech, music, and general audio. Code for reproducing the results and supporting webpage are available at https://github.com/gnvIRCAM/Audio-representation-trimming

Current generative models are able to generate high-quality artefacts but have been shown to struggle with compositional reasoning, which can be defined as the ability to generate complex structures from simpler elements. In this paper, we focus on the problem of compositional representation learning for music data, specifically targeting the fully-unsupervised setting. We propose a simple and extensible framework that leverages an explicit compositional inductive bias, defined by a flexible auto-encoding objective that can leverage any of the current state-of-art generative models. We demonstrate that our framework, used with diffusion models, naturally addresses the task of unsupervised audio source separation, showing that our model is able to perform high-quality separation. Our findings reveal that our proposal achieves comparable or superior performance with respect to other blind source separation methods and, furthermore, it even surpasses current state-of-art supervised baselines on signal-to-interference ratio metrics. Additionally, by learning an a-posteriori masking diffusion model in the space of composable representations, we achieve a system capable of seamlessly performing unsupervised source separation, unconditional generation, and variation generation. Finally, as our proposal works in the latent space of pre-trained neural audio codecs, it also provides a lower computational cost with respect to other neural baselines.

Deep generative models are now able to synthesize high-quality audio signals, shifting the critical aspect in their development from audio quality to control capabilities. Although text-to-music generation is getting largely adopted by the general public, explicit control and example-based style transfer are more adequate modalities to capture the intents of artists and musicians. In this paper, we aim to unify explicit control and style transfer within a single model by separating local and global information to capture musical structure and timbre respectively. To do so, we leverage the capabilities of diffusion autoencoders to extract semantic features, in order to build two representation spaces. We enforce disentanglement between those spaces using an adversarial criterion and a two-stage training strategy. Our resulting model can generate audio matching a timbre target, while specifying structure either with explicit controls or through another audio example. We evaluate our model on one-shot timbre transfer and MIDI-to-audio tasks on instrumental recordings and show that we outperform existing baselines in terms of audio quality and target fidelity. Furthermore, we show that our method can generate cover versions of complete musical pieces by transferring rhythmic and melodic content to the style of a target audio in a different genre.

In recent years, significant advances have been made in deep learning models for audio generation, offering promising tools for musical creation. In this work, we investigate the use of deep audio generative models in interactive dance/music performance. We adopted a performance-led research design approach, establishing an art-research collaboration between a researcher/musician and a dancer. First, we describe our motion-sound interactive system integrating deep audio generative model and propose three methods for embodied exploration of deep latent spaces. Then, we detail the creative process for building the performance centered on the co-design of the system. Finally, we report feedback from the dancer’s interviews and discuss the results and perspectives. The code implementation is publicly available on our github1.

Machine learning approaches now achieve impressive generation capabilities in numerous domains such as image, audio or video. However, most training \& evaluation frameworks revolve around the idea of strictly modelling the original data distribution rather than trying to extrapolate from it. This precludes the ability of such models to diverge from the original distribution and, hence, exhibit some creative traits. In this paper, we propose various perspectives on how this complicated goal could ever be achieved, and provide preliminary results on our novel training objective called \textit{Bounded Adversarial Divergence} (BAD).

The development of generative Machine Learning (ML) models in creative practices, enabled by the recent improvements in usability and availability of pre-trained models, is raising more and more interest among artists, practitioners and performers. Yet, the introduction of such techniques in artistic domains also revealed multiple limitations that escape current evaluation methods used by scientists. Notably, most models are still unable to generate content that lay outside of the domain defined by the training dataset. In this paper, we propose an alternative prospective framework, starting from a new general formulation of ML objectives, that we derive to delineate possible implications and solutions that already exist in the ML literature (notably for the audio and musical domain). We also discuss existing relations between generative models and computational creativity and how our framework could help address the lack of creativity in existing models.

Deep generative models applied to audio have improved by a large margin the state-of-the-art in many speech and music related tasks. However, as raw waveform modelling remains an inherently difficult task, audio generative models are either computationally intensive, rely on low sampling rates, are complicated to control or restrict the nature of possible signals. Among those models, Variational AutoEncoders (VAE) give control over the generation by exposing latent variables, although they usually suffer from low synthesis quality. In this paper, we introduce a Realtime Audio Variational autoEncoder (RAVE) allowing both fast and high-quality audio waveform synthesis. We introduce a novel two-stage training procedure, namely representation learning and adversarial fine-tuning. We show that using a post-training analysis of the latent space allows a direct control between the reconstruction fidelity and the representation compactness. By leveraging a multi-band decomposition of the raw waveform, we show that our model is the first able to generate 48kHz audio signals, while simultaneously running 20 times faster than real-time on a standard laptop CPU. We evaluate synthesis quality using both quantitative and qualitative subjective experiments and show the superiority of our approach compared to existing models. Finally, we present applications of our model for timbre transfer and signal compression. All of our source code and audio examples are publicly available.

The problem of target-based computer-aided orchestration is a recurring topic in the contemporary music community. Because of its complexity, computer-aided orchestration remains a partially unsolved problem and several systems have been developed in the last twenty years. This article presents a practical overview of the recently introduced Orchidea framework for dynamic computer-aided target-based orchestration. Orchidea continues the line of tools dedicated to the subject (the so-called Orchid* family) originally developed at the Institut de Recherche et Coordination Acoustique/Musique in Paris. Unlike its predecessors, Orchidea uses a combination of optimization techniques that include stochastic matching pursuit, long short-term memory neural networks, and monoobjective evolutionary optimization, with a specifically designed cost function. Symbolic constraints can be integrated in the cost function, and temporally evolving sounds are handled by segmenting them into a set of static targets optimized jointly and then connected. Orchidea is deployed in three different ways: a standalone application, designed to streamline a simplified compositional workflow; a Max package, targeted at composers willing to connect target-based orchestration to the more general area of computer-aided composition; and a set of command-line tools, mostly intended for research purposes and batch processing. The main aim of this article is to present an overview of such software systems and show several instances of the Orchidea framework's application in recent musical productions, tracing the path for future research on the subject.

Recently, the field of musical co-creativity has gained some momentum. In this context, our goal is twofold: to develop an intelligent listening and predictive module of chord sequences, and to propose an adapted evaluation of the associated Music Information Retrieval (MIR) tasks that are the real-time extraction of musical chord labels from a live audio stream and the prediction of a possible continuation of the extracted symbolic sequence. Indeed, this application case invites us to raise questions about the evaluation processes and methodology that are currently applied to chord-based MIR models. In this paper, we focus on musical chords since these mid-level features are frequently used to describe harmonic progressions in Western music. In the case of chords, there exists some strong inherent hierarchical and functional relationships. However, most of the research in the field of MIR focuses mainly on the performance of chord-based statistical models, without considering music-based evaluation or learning. Indeed, usual evaluations are based on a binary qualification of the classification outputs (right chord predicted versus wrong chord predicted). Therefore, we present a specifically-tailored chord analyser to measure the performances of chord-based models in terms of functional qualification of the classification outputs (by taking into account the harmonic function of the chords). Then, in order to introduce musical knowledge into the learning process for the automatic chord extraction task, we propose a specific musical distance for comparing predicted and labeled chords. Finally, we conduct investigations into the impact of including high-level metadata in chord sequence prediction learning (such as information on key or downbeat position). We show that a model can obtain better performances in terms of accuracy or perplexity, but output biased results. At the same time, a model with a lower accuracy score can output errors with more musical meaning. Therefore, performing a goal-oriented evaluation allows a better understanding of the results and a more adapted design of MIR models.

Granular sound synthesis is a popular audio generation technique based on rearranging sequences of small waveform windows. In order to control the synthesis, all grains in a given corpus are analyzed through a set of acoustic descriptors. This provides a representation reflecting some form of local similarities across the grains. However, the quality of this grain space is bound by that of the descriptors. Its traversal is not continuously invertible to signal and does not render any structured temporality. We demonstrate that generative neural networks can implement granular synthesis while alleviating most of its shortcomings. We efficiently replace its audio descriptor basis by a probabilistic latent space learned with a Variational Auto-Encoder. In this setting the learned grain space is invertible, meaning that we can continuously synthesize sound when traversing its dimensions. It also implies that original grains are not stored for synthesis. Another major advantage of our approach is to learn structured paths inside this latent space by training a higher-level temporal embedding over arranged grain sequences. The model can be applied to many types of libraries, including pitched notes or unpitched drums and environmental noises. We report experiments on the common granular synthesis processes as well as novel ones such as conditional sampling and morphing.