Eurasip Journal on Audio Speech and Music Processing最新文献

In the development of acoustic signal processing algorithms, their evaluation in various acoustic environments is of utmost importance. In order to advance evaluation in realistic and reproducible scenarios, several high-quality acoustic databases have been developed over the years. In this paper, we present another complementary database of acoustic recordings, referred to as the Multi-arraY Room Acoustic Database (MYRiAD). The MYRiAD database is unique in its diversity of microphone configurations suiting a wide range of enhancement and reproduction applications (such as assistive hearing, teleconferencing, or sound zoning), the acoustics of the two recording spaces, and the variety of contained signals including 1214 room impulse responses (RIRs), reproduced speech, music, and stationary noise, as well as recordings of live cocktail parties held in both rooms. The microphone configurations comprise a dummy head (DH) with in-ear omnidirectional microphones, two behind-the-ear (BTE) pieces equipped with 2 omnidirectional microphones each, 5 external omnidirectional microphones (XMs), and two concentric circular microphone arrays (CMAs) consisting of 12 omnidirectional microphones in total. The two recording spaces, namely the SONORA Audio Laboratory (SAL) and the Alamire Interactive Laboratory (AIL), have reverberation times of 2.1 s and 0.5 s, respectively. Audio signals were reproduced using 10 movable loudspeakers in the SAL and a built-in array of 24 loudspeakers in the AIL. MATLAB and Python scripts are included for accessing the signals as well as microphone and loudspeaker coordinates. The database is publicly available (https://zenodo.org/record/7389996).

The human auditory system employs a number of principles to facilitate the selection of perceptually separated streams from a complex sound mixture. The brain leverages multi-scale redundant representations of the input and uses memory (or priors) to guide the selection of a target sound from the input mixture. Moreover, feedback mechanisms refine the memory constructs resulting in further improvement of selectivity of a particular sound object amidst dynamic backgrounds. The present study proposes a unified end-to-end computational framework that mimics these principles for sound source separation applied to both speech and music mixtures. While the problems of speech enhancement and music separation have often been tackled separately due to constraints and specificities of each signal domain, the current work posits that common principles for sound source separation are domain-agnostic. In the proposed scheme, parallel and hierarchical convolutional paths map input mixtures onto redundant but distributed higher-dimensional subspaces and utilize the concept of temporal coherence to gate the selection of embeddings belonging to a target stream abstracted in memory. These explicit memories are further refined through self-feedback from incoming observations in order to improve the system's selectivity when faced with unknown backgrounds. The model yields stable outcomes of source separation for both speech and music mixtures and demonstrates benefits of explicit memory as a powerful representation of priors that guide information selection from complex inputs.

Instrumentalplaying techniques such as vibratos, glissandos, and trills often denote musical expressivity, both in classical and folk contexts. However, most existing approaches to music similarity retrieval fail to describe timbre beyond the so-called "ordinary" technique, use instrument identity as a proxy for timbre quality, and do not allow for customization to the perceptual idiosyncrasies of a new subject. In this article, we ask 31 human participants to organize 78 isolated notes into a set of timbre clusters. Analyzing their responses suggests that timbre perception operates within a more flexible taxonomy than those provided by instruments or playing techniques alone. In addition, we propose a machine listening model to recover the cluster graph of auditory similarities across instruments, mutes, and techniques. Our model relies on joint time-frequency scattering features to extract spectrotemporal modulations as acoustic features. Furthermore, it minimizes triplet loss in the cluster graph by means of the large-margin nearest neighbor (LMNN) metric learning algorithm. Over a dataset of 9346 isolated notes, we report a state-of-the-art average precision at rank five (AP@5) of 99.0%±1. An ablation study demonstrates that removing either the joint time-frequency scattering transform or the metric learning algorithm noticeably degrades performance.

Amongst the various characteristics of a speech signal, the expression of emotion is one of the characteristics that exhibits the slowest temporal dynamics. Hence, a performant speech emotion recognition (SER) system requires a predictive model that is capable of learning sufficiently long temporal dependencies in the analysed speech signal. Therefore, in this work, we propose a novel end-to-end neural network architecture based on the concept of dilated causal convolution with context stacking. Firstly, the proposed model consists only of parallelisable layers and is hence suitable for parallel processing, while avoiding the inherent lack of parallelisability occurring with recurrent neural network (RNN) layers. Secondly, the design of a dedicated dilated causal convolution block allows the model to have a receptive field as large as the input sequence length, while maintaining a reasonably low computational cost. Thirdly, by introducing a context stacking structure, the proposed model is capable of exploiting long-term temporal dependencies hence providing an alternative to the use of RNN layers. We evaluate the proposed model in SER regression and classification tasks and provide a comparison with a state-of-the-art end-to-end SER model. Experimental results indicate that the proposed model requires only 1/3 of the number of model parameters used in the state-of-the-art model, while also significantly improving SER performance. Further experiments are reported to understand the impact of using various types of input representations (i.e. raw audio samples vs log mel-spectrograms) and to illustrate the benefits of an end-to-end approach over the use of hand-crafted audio features. Moreover, we show that the proposed model can efficiently learn intermediate embeddings preserving speech emotion information.

If music is the language of the universe, musical note onsets may be the syllables for this language. Not only do note onsets define the temporal pattern of a musical piece, but their time-frequency characteristics also contain rich information about the identity of the musical instrument producing the notes. Note onset detection (NOD) is the basic component for many music information retrieval tasks and has attracted significant interest in audio signal processing research. In this paper, we propose an NOD method based on a novel feature coined as Normalized Identification of Note Onset based on Spectral Sparsity (NINOS²). The NINOS² feature can be thought of as a spectral sparsity measure, aiming to exploit the difference in spectral sparsity between the different parts of a musical note. This spectral structure is revealed when focusing on low-magnitude spectral components that are traditionally filtered out when computing note onset features. We present an extensive set of NOD simulation results covering a wide range of instruments, playing styles, and mixing options. The proposed algorithm consistently outperforms the baseline Logarithmic Spectral Flux (LSF) feature for the most difficult group of instruments which are the sustained-strings instruments. It also shows better performance for challenging scenarios including polyphonic music and vibrato performances.

Over the recent years, machine learning techniques have been employed to produce state-of-the-art results in several audio related tasks. The success of these approaches has been largely due to access to large amounts of open-source datasets and enhancement of computational resources. However, a shortcoming of these methods is that they often fail to generalize well to tasks from real life scenarios, due to domain mismatch. One such task is foreground speech detection from wearable audio devices. Several interfering factors such as dynamically varying environmental conditions, including background speakers, TV, or radio audio, render foreground speech detection to be a challenging task. Moreover, obtaining precise moment-to-moment annotations of audio streams for analysis and model training is also time-consuming and costly. In this work, we use multiple instance learning (MIL) to facilitate development of such models using annotations available at a lower time-resolution (coarsely labeled). We show how MIL can be applied to localize foreground speech in coarsely labeled audio and show both bag-level and instance-level results. We also study different pooling methods and how they can be adapted to densely distributed events as observed in our application. Finally, we show improvements using speech activity detection embeddings as features for foreground detection.