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.

Despite significant advances in deep models for music generation, the use of these techniques remains restricted to expert users. Before being democratized among musicians, generative models must first provide expressive control over the generation, as this conditions the integration of deep generative models in creative workflows. In this paper, we tackle this issue by introducing a deep generative audio model providing expressive and continuous descriptor-based control, while remaining lightweight enough to be embedded in a hardware synthesizer. We enforce the controllability of real-time generation by explicitly removing salient musical features in the latent space using an adversarial confusion criterion. User-specified features are then reintroduced as additional conditioning information, allowing for continuous control of the generation, akin to a synthesizer knob. We assess the performance of our method on a wide variety of sounds including instrumental, percussive and speech recordings while providing both timbre and attributes transfer, allowing new ways of generating sounds.

Deep learning models are now core components of modern audio synthesis, and their use has increased significantly in recent years, leading to highly accurate systems for multiple tasks. However, this quest for quality comes at a tremendous computational cost, which incurs vast energy consumption and greenhouse gas emissions. At the heart of this problem are the standardized evaluation metrics used by the scientific community to compare various contributions. In this paper, we suggest relying on a multi-objective metric based on Pareto optimality, which considers equally the accuracy and energy consumption of a model. By applying our measure to the current state-of-the-art in generative audio models, we show that it can drastically change the significance of the results. We hope to raise awareness on the need to more systematically investigate the energy efficiency of high-quality models, in order to place computational costs at the center of deep learning research priorities.

Audio synthesizers are pervasive in modern music production. These highly complex audio generation functions provide a unique diversity through their large sets of parameters. However, this feature also can make them extremely hard and obfuscated to use, especially for non-expert users with no formal knowledge on signal processing. We recently introduced a novel formalization of the problem of synthesizer control as learning an invertible mapping between an audio latent space, extracted from the audio signal, and a target parameter latent space, extracted from the synthesizer's presets, using normalizing flows. In addition to model a continuous representation allowing to ease the intuitive exploration of the synthesizer, it also provides a ground-breaking method for audio-based parameter inference, vocal control and macro-control learning. Here, we discuss the details of integrating these high-level features to develop new interaction schemes between a human user and the generating device: parameters inference from audio, high-level preset visualization and interpolation, that can be used both in off-time and real-time situations. Moreover, we also leverage LeapMotion devices to allow the control of hundreds of parameters simply by moving one hand across space to explore the low-dimensional latent space, allowing to both empower and facilitate the user's interaction with the synthesizer.

Seeking sound samples in a massive database can be a tedious and time consuming task. Even when metadata are available, query results may remain far from the timbre expected by users. This problem stems from the nature of query specification, which does not account for the underlying complexity of audio data. The Query By Example (QBE) paradigm tries to tackle this shortcoming by finding audio clips similar to a given sound example. However, it requires users to have a well-formed soundfile of what they seek, which is not always a valid assumption. Furthermore, most audio-retrieval systems rely on a single measure of similarity, which is unlikely to convey the perceptual similarity of audio signals. We address in this paper an innovative way of querying generic audio databases by simultaneously optimizing the temporal evolution of multiple spectral properties. We show how this problem can be cast into a new approach merging multiobjective optimization and time series matching, called MultiObjective Time Series (MOTS) matching. We formally state this problem and report an efficient implementation. This approach introduces a multidimensional assessment of similarity in audio matching. This allows to cope with the multidimensional nature of timbre perception and also to obtain a set of efficient propositions rather than a single best solution. To demonstrate the performances of our approach, we show its efficiency in audio classification tasks. By introducing a selection criterion based on the hypervolume dominated by a class, we show that our approach outstands the state-of-art methods in audio classification even with a few number of features. We demonstrate its robustness to several classes of audio distortions. Finally, we introduce two innovative applications of our method for sound querying.