Recent advancements in deep learning have led to widespread use of techniques for audio content generation, notably employing Denoising Diffusion Probabilistic Models (DDPM) across various tasks. Among these, Foley Sound Synthesis is of particular interest for its role in applications for the creation of multimedia content. Given the temporal-dependent nature of sound, it is crucial to design generative models that can effectively handle the sequential modeling of audio samples. Selective State-Space Models (SSMs) have recently been proposed as a valid alternative to previously proposed techniques, demonstrating competitive performance with lower computational complexity. In this paper, we introduce MambaFoley, a diffusion-based model that, to the best of our knowledge, is the first to leverage the recently proposed SSM known as Mamba for the Foley sound generation task. To evaluate the effectiveness of the proposed method, we compare it with a state-of-the-art Foley sound generative model using both objective and subjective analyses.

Head-Related Transfer Functions (HRTFs) have fundamental applications for realistic rendering in immersive audio scenarios. However, they are strongly subject-dependent as they vary considerably depending on the shape of the ears, head and torso. Thus, personalization procedures are required for accurate binaural rendering. Recently, Denoising Diffusion Probabilistic Models (DDPMs), a class of generative learning techniques, have been applied to solve a variety of signal processing-related problems. In this paper, we propose a first approach for using DDPM conditioned on anthropometric measurements to generate personalized Head-Related Impulse Response (HRIR), the time-domain representation of HRTF. The results show the feasibility of DDPMs for HRTF personalization obtaining performance in line with state-of-the-art models.

Reconstructing the sound field in a room is an important task for several applications, such as sound control and augmented (AR) or virtual reality (VR). In this paper, we propose a data-driven generative model for reconstructing the magnitude of acoustic fields in rooms with a focus on the modal frequency range. We introduce, for the first time, the use of a conditional Denoising Diffusion Probabilistic Model (DDPM) trained in order to reconstruct the sound field (SF-Diff) over an extended domain. The architecture is devised in order to be conditioned on a set of limited available measurements at different frequencies and generate the sound field in target, unknown, locations. The results show that SF-Diff is able to provide accurate reconstructions. We conduct a comparative analysis with two state-of-the-art baseline methods, one relying on kernel interpolation and the other on deep learning.

In this manuscript, we propose a novel method to perform audio inpainting, i.e., the restoration of audio signals presenting multiple missing parts. Audio inpainting can be interpreted in the context of inverse problems as the task of reconstructing an audio signal from its corrupted observation. For this reason, our method is based on a deep prior approach, a recently proposed technique that proved to be effective in the solution of many inverse problems, among which image inpainting. Deep prior allows one to consider the structure of a neural network as an implicit prior and to adopt it as a regularizer. Differently from the classical deep learning paradigm, deep prior performs a single-element training and thus it can be applied to corrupted audio signals independently from the available training data sets. In the context of audio inpainting, a network presenting relevant audio priors will possibly generate a restored version of an audio signal, only provided with its corrupted observation. Our method exploits a time-frequency representation of audio signals and makes use of a multi-resolution convolutional autoencoder, that has been enhanced to perform the harmonic convolution operation. Results show that the proposed technique is able to provide a coherent and meaningful reconstruction of the corrupted audio. It is also able to outperform the methods considered for comparison, in its domain of application.

The problems of speech separation and enhancement concern the extraction of the speech emitted by a target speaker when placed in a scenario where multiple interfering speakers or noise are present, respectively. A plethora of practical applications such as home assistants and teleconferencing require some sort of speech separation and enhancement pre-processing before applying Automatic Speech Recognition (ASR) systems. In the recent years, most techniques have focused on the application of deep learning to either time-frequency or time-domain representations of the input audio signals. In this paper we propose a real-time multichannel speech separation and enhancement technique, which is based on the combination of a directional representation of the sound field, denoted as beamspace, with a lightweight Convolutional Neural Network (CNN). We consider the case where the Direction-Of-Arrival (DOA) of the target speaker is approximately known, a scenario where the power of the beamspace-based representation can be fully exploited, while we make no assumption regarding the identity of the talker. We present experiments where the model is trained on simulated data and tested on real recordings and we compare the proposed method with a similar state-of-the-art technique.

The complete understanding of the decision-making process of Convolutional Neural Networks (CNNs) is far from being fully reached. Many researchers proposed techniques to interpret what a network actually “learns” from data. Nevertheless many questions still remain unanswered. In this work we study one aspect of this problem by reconstructing speech from the intermediate embeddings computed by a CNNs. Specifically, we consider a pre-trained network that acts as a feature extractor from speech audio. We investigate the possibility of inverting these features, reconstructing the input signals in a black-box scenario, and quantitatively measure the reconstruction quality by measuring the word-error-rate of an off-the-shelf ASR model. Experiments performed using two different CNN architectures trained for six different classification tasks, show that it is possible to reconstruct time-domain speech signals that preserve the semantic content, whenever the embeddings are extracted before the fully connected layers.

The interest in deep learning methods for solving traditional signal processing tasks has been steadily growing in the last years. Time delay estimation (TDE) in adverse scenarios is a challenging problem, where classical approaches based on generalized cross-correlations (GCCs) have been widely used for decades. Recently, the frequency-sliding GCC (FS-GCC) was proposed as a novel technique for TDE based on a sub-band analysis of the cross-power spectrum phase, providing a structured two-dimensional representation of the time delay information contained across different frequency bands. Inspired by deep-learning-based image denoising solutions, we propose in this paper the use of convolutional neural networks (CNNs) to learn the time-delay patterns contained in FS-GCCs extracted in adverse acoustic conditions. Our experiments confirm that the proposed approach provides excellent TDE performance while being able to generalize to different room and sensor setups.

The generalized cross-correlation(GCC) is regarded as the most popular approach for estimating the time difference of arrival (TDOA) between the signals received at two sensors. Time delay estimates are obtained by maximizing the GCC output, where the direct-path delay is usually observed as a prominent peak. Moreover, GCCs play also an important role in steered response power (SRP) localization algorithms, where the SRP functional can be written as an accumulation of the GCCs computed from multiple sensor pairs. Unfortunately, the accuracy of TDOA estimates is affected by multiple factors, including noise, reverberation and signal bandwidth. In this paper, a sub-band approach for time delay estimation aimed at improving the performance of the conventional GCC is presented. The proposed method is based on the extraction of multiple GCCs corresponding to different frequency bands of the cross-power spectrum phase in a sliding-window fashion. The major contributions of this paper include: 1) a sub-band GCC representation of the cross-power spectrum phase that, despite having a reduced temporal resolution, provides a more suitable representation for estimating the true TDOA; 2) such matrix representation is shown to be rank one in the ideal noiseless case, a property that is exploited in more adverse scenarios to obtain a more robust and accurate GCC; 3) we propose a set of low-rank approximation alternatives for processing the sub-band GCC matrix, leading to better TDOA estimates and source localization performance. An extensive set of experiments is presented to demonstrate the validity of the proposed approach.

In this article we present a methodology for source localization in reverberant environments from Generalized Cross Correlations (GCCs) computed between spatially distributed individual microphones. Reverberation tends to negatively affect localization based on Time Differences of Arrival (TDOAs), which become inaccurate due to the presence of spurious peaks in the GCC. We therefore adopt a data-driven approach based on a convolutional neural network, which, using the GCCs as input, estimates the source location in two steps. It first computes the Ray Space Transform (RST) from multiple arrays. The RST is a convenient representation of the acoustic rays impinging on the array in a parametric space, called Ray Space. Rays produced by a source are visualized in the RST as patterns, whose position is uniquely related to the source location. The second step consists of estimating the source location through a nonlinear fitting, which estimates the coordinates that best approximate the RST pattern obtained through the first step. It is worth noting that training can be accomplished on simulated data only, thus relaxing the need of actually deploying microphone arrays in the acoustic scene. The localization accuracy of the proposed techniques is similar to the one of SRP-PHAT, however our method demonstrates an increased robustness regarding different distributed microphones configurations. Moreover, the use of the RST as an intermediate representation makes it possible for the network to generalize to data unseen during training.

Cooperative music making in networked environments has been subject of extensive research, scientific and artistic. Networked music performance (NMP) is attracting renewed interest thanks to the growing availability of effective technology and tools for computer-based communications, especially in the area of distance and blended learning applications. We propose a conceptual framework for NMP research and design in the context of classical chamber music practice and learning: presence-related constructs and objective quality metrics are used to problematize and systematize the many factors affecting the experience of studying and practicing music in a networked environment. To this end, a preliminary NMP experiment on the effect of latency on chamber music duos experience and quality of the performance is introduced. The degree of involvement, perceived coherence, and immersion of the NMP environment are here combined with measures on the networked performance, including tempo trends and misalignments from the shared score. Early results on the impact of temporal factors on NMP musical interaction are outlined, and their methodological implications for the design of pedagogical applications are discussed.