Paper Archive

Recent advancements in deep learning have shown great potential for audio applications, improving the accuracy of previous solutions for tasks such as music transcription, beat detection, and real-time audio processing. In addition, the availability of increasingly powerful embedded computers has led many deep learning framework developers to devise software optimized to run pretrained models in resource-constrained contexts. As a result, the use of deep learning on embedded devices and audio plugins has become more widespread. However, confusion has been rising around deep learning inference engines, regarding which of these can run in real-time and which are less resource-hungry. In this paper, we present a comparison of four available deep learning inference engines for real-time audio classification on the CPU of an embedded single-board computer: TensorFlow Lite, TorchScript, ONNX Runtime, and RTNeural. Results show that all inference engines can execute neural network models in real-time with appropriate code practices, but execution time varies between engines and models. Most importantly, we found that most of the less-specialized engines offer great flexibility and can be used effectively for real-time audio classification, with slightly better results than a real-time-specific approach. In contrast, more specialized solutions can offer a lightweight and minimalist alternative where less flexibility is needed.

Download

Deep learning models applied to raw audio are rapidly gaining relevance in modeling audio analog devices. This paper investigates the use of different deep architectures for modeling audio optical compression. The models use as input and produce as output raw audio samples at audio rate, and it works with noor small-input buffers allowing a theoretical real-time and lowlatency implementation. In this study, two compressor parameters, the ratio, and threshold have been included in the modeling process aiming to condition the inference of the trained network. Deep learning architectures are compared to model an all-tube optical mono compressor including feed-forward, recurrent, and encoder-decoder models. The results of this study show that feedforward and long short-term memory architectures present limitations in modeling the triggering phase of the compressor, performing well only on the sustained phase. On the other hand, encoderdecoder models outperform other architectures in replicating the overall compression process, but they overpredict the energy of high-frequency components.

Download

Deep learning models have provided extremely successful solutions in most audio application fields. However, the high accuracy of these models comes at the expense of a tremendous computation cost. This aspect is almost always overlooked in evaluating the quality of proposed models. However, models should not be evaluated without taking into account their complexity. This aspect is especially critical in audio applications, which heavily relies on specialized embedded hardware with real-time constraints. In this paper, we build on recent observations that deep models are highly overparameterized, by studying the lottery ticket hypothesis on deep generative audio models. This hypothesis states that extremely efficient small sub-networks exist in deep models and would provide higher accuracy than larger models if trained in isolation. However, lottery tickets are found by relying on unstructured masking, which means that resulting models do not provide any gain in either disk size or inference time. Instead, we develop here a method aimed at performing structured trimming. We show that this requires to rely on global selection and introduce a specific criterion based on mutual information. First, we confirm the surprising result that smaller models provide higher accuracy than their large counterparts. We further show that we can remove up to 95% of the model weights without significant degradation in accuracy. Hence, we can obtain very light models for generative audio across popular methods such as Wavenet, SING or DDSP, that are up to 100 times smaller with commensurate accuracy. We study the theoretical bounds for embedding these models on Raspberry Pi and Arduino, and show that we can obtain generative models on CPU with equivalent quality as large GPU models. Finally, we discuss the possibility of implementing deep generative audio models on embedded platforms.

Download

Identifying instrument activities within audio excerpts is vital in music information retrieval, with significant implications for music cataloging and discovery. Prior deep learning endeavors in musical instrument recognition have predominantly emphasized instrument classes with ample data availability. Recent studies have demonstrated the applicability of hierarchical classification in detecting instrument activities in orchestral music, even with limited fine-grained annotations at the instrument level. Based on the Hornbostel-Sachs classification, such a hierarchical classification system is evaluated using the MedleyDB dataset, renowned for its diversity and richness concerning various instruments and music genres. This work presents various strategies to integrate hierarchical structures into models and tests a new class of models for hierarchical music prediction. This study showcases more reliable coarse-level instrument detection by bridging the gap between detailed instrument identification and group-level recognition, paving the way for further advancements in this domain.

Download

Controllable timbre synthesis has been a subject of research for several decades, and deep neural networks have been the most successful in this area. Deep generative models such as Variational Autoencoders (VAEs) have the ability to generate a high-level representation of audio while providing a structured latent space. Despite their advantages, the interpretability of these latent spaces in terms of human perception is often limited. To address this limitation and enhance the control over timbre generation, we propose a regularized VAE-based latent space that incorporates timbre descriptors. Moreover, we suggest a more concise representation of sound by utilizing its harmonic content, in order to minimize the dimensionality of the latent space.

Download

The noise that lavalier microphones produce when rubbing against clothing (typically referred to as rustle) can be extremely difficult to automatically remove because it is highly non-stationary and overlaps with speech in both time and frequency. Recent breakthroughs in deep neural networks have led to novel techniques for separating speech from non-stationary background noise. In this paper, we apply neural network speech separation techniques to remove rustle noise, and quantitatively compare multiple deep network architectures and input spectral resolutions. We find the best performance using bidirectional recurrent networks and spectral resolution of around 20 Hz. Furthermore, we propose an ambience preservation post-processing step to minimize potential gating artifacts during pauses in speech.

Download

One of the most challenging tasks in physically-informed sound synthesis is the estimation of model parameters to produce a desired timbre. Automatic parameter estimation procedures have been developed in the past for some specific parameters or application scenarios but, up to now, no approach has been proved applicable to a wide variety of use cases. A general solution to parameters estimation problem is provided along this paper which is based on a supervised convolutional machine learning paradigm. The described approach can be classified as “end-to-end” and requires, thus, no specific knowledge of the model itself. Furthermore, parameters are learned from data generated by the model, requiring no effort in the preparation and labeling of the training dataset. To provide a qualitative and quantitative analysis of the performance, this method is applied to a patented digital waveguide pipe organ model, yielding very promising results.

Download

Loudspeaker equalization is an established topic in the literature, and currently many techniques are available to address most practical use cases. However, most of these rely on accurate measurements of the loudspeaker in an anechoic environment, which in some occurrences is not feasible. This is the case, e.g. of custom digital organs, which have a set of loudspeakers that are built into a large and geometrically-complex piece of furniture, which may be too heavy and large to be transported to a measurement room, or may require a big one, making traditional impulse response measurements impractical for most users. In this work we propose a method to find the inverse of the sound emission system in a reverberant environment, based on a Deep Learning dereverberation algorithm. The method is agnostic of the room characteristics and can be, thus, conducted in an automated fashion in any environment. A real use case is discussed and results are provided, showing the effectiveness of the approach in designing filters that match closely the magnitude response of the ideal inverting filters.

Download

This study investigates the use of an unsupervised, physicsinformed deep learning framework to model a one-degree-offreedom mass-spring system subjected to a nonlinear friction bow force and governed by a set of ordinary differential equations. Specifically, it examines the application of Physics-Informed Neural Networks (PINNs) and Physics-Informed Deep Operator Networks (PI-DeepONets). Our findings demonstrate that PINNs successfully address the problem across different bow force scenarios, while PI-DeepONets perform well under low bow forces but encounter difficulties at higher forces. Additionally, we analyze the Hessian eigenvalue density and visualize the loss landscape. Overall, the presence of large Hessian eigenvalues and sharp minima indicates highly ill-conditioned optimization. These results underscore the promise of physics-informed deep learning for nonlinear modelling in musical acoustics, while also revealing the limitations of relying solely on physics-based approaches to capture complex nonlinearities. We demonstrate that PI-DeepONets, with their ability to generalize across varying parameters, are well-suited for sound synthesis. Furthermore, we demonstrate that the limitations of PI-DeepONets under higher forces can be mitigated by integrating observation data within a hybrid supervised-unsupervised framework. This suggests that a hybrid supervised-unsupervised DeepONets framework could be a promising direction for future practical applications.

Download

Audio processors whose parameters are modified periodically over time are often referred as time-varying or modulation based audio effects. Most existing methods for modeling these type of effect units are often optimized to a very specific circuit and cannot be efficiently generalized to other time-varying effects. Based on convolutional and recurrent neural networks, we propose a deep learning architecture for generic black-box modeling of audio processors with long-term memory. We explore the capabilities of deep neural networks to learn such long temporal dependencies and we show the network modeling various linear and nonlinear, time-varying and time-invariant audio effects. In order to measure the performance of the model, we propose an objective metric based on the psychoacoustics of modulation frequency perception. We also analyze what the model is actually learning and how the given task is accomplished.

Download