Search

State-of-the-art algorithms in many music information retrieval (MIR) tasks such as chord recognition, multipitch estimation, or instrument recognition rely on deep learning algorithms, which require large amounts of data to be trained and evaluated. In this paper, we present the IDMT-SMT-CHORD-SEQUENCES dataset, which is a novel synthetic dataset of 15,000 chord progressions played on 45 different musical instruments. The dataset is organized in a triplet fashion and each triplet includes one "anchor" chord sequence as well as one corresponding similar and dissimilar chord progression. The audio files are synthesized from MIDI data using FluidSynth with a selected sound font. Furthermore, we conducted a benchmark experiment on time-dependent harmonic similarity based on learnt embedding representations. The results show that a convolutional neural network (CNN), which considers the temporal context of a chord progression, outperforms a simpler approach based on temporal averaging of input features.

Selective Hearing (SH) refers to the listener’s attention to specific sound sources of interest in their auditory scene. Achieving SH through computational means involves detection, classification, separation, localization and enhancement of sound sources. Deep neural networks (DNNs) have been shown to perform these tasks in a robust and time-efficient manner. A promising application of SH are intelligent noise-cancelling headphones, where sound sources of interest, such as warning signals, sirens or speech, are extracted from a given auditory scene and conveyed to the user, whilst the rest of the auditory scene remains inaudible. For this purpose, existing noise cancellation approaches need to be combined with machine learning techniques. In this context, we evaluate a convolutional neural network (CNN) architecture and a long short-term memory (LSTM) architecture for the detection and separation of sirens. In addition, we propose a data simulation approach for generating different sound environments for a virtual pair of headphone microphones. The Fraunhofer SpatialSound Wave technology is used for a realistic evaluation of the trained models. For the evaluation, a three-dimensional acoustic scene is simulated via the object-based audio approach.

Audio-Forensik, Partielles Audio-Matching und Audio-Phylogenie-Analyse sind drei sehr spezialisierte Technologien, die jedoch in einer Vielzahl von Anwendungen wie Manipulationserkennung, Duplikaterkennung, Rechtemanagement, Qualitätskontrolle und Programmanalyse eingesetzt werden können. Der Artikel beschreibt die entsprechenden Ansätze und erklärt, wie sie - insbesondere in Kombination - zur Lösung sehr unterschiedlicher Probleme eingesetzt werden können.

Smart city applications for acoustic monitoring become essential to cope with the overall increasing noise pollution in urban environments. This paper gives an overview over a distributed sensor network for noise monitoring in the German city of Jena. Several acoustic sensor units allow for classifying among various acoustic scenes and events using Convolutional Neural Networks (CNN) and measuring different noise level parameters. Connected by a communication system based on MQTT (Message Queue Telemetry Transport), these sensors communicate measurement data to a central server for data postprocessing and storage. Finally, a web-based application allows for various real-time visualizations of noise exposure distributed over the city.

Automatic speech/music discrimination describes the task of automatically detecting speech and music audio within a recording. This is useful for a great number of tasks in both research and industry. In particular, this approach can be used for broadcast signals (e.g. from TV or radio stations) in order to determine the amount of music played. The results can then be used for various reporting purposes (e.g. for royalty collection societies such as the German GEMA). Speech/music discrimination is commonly performed by using machine learning technologies, where models are first trained on manually annotated data, and can then be used to classify previously unseen audio data. In this paper, we give an overview over the applications and the state of the art of speech/music discrimination. Afterwards, we present our approaches based on a set of audio features, Gaussian Mixture Models and Deep Learning. Finally, we give suggestions for the direction of new research into this topic.

1. Passive acoustic monitoring is frequently used for marine mammals, and more recently it has also become popular for terrestrial species. Key advantages are the monitoring of (1) elusive species, (2) different taxa simultaneously, (3) large temporal and spatial scales, (4) with reduced human presence and (5) with considerable time saving for data processing. However, terrestrial sound environments can be highly complex; they are very challenging when trying to automatically detect and classify vocalizations because of low signal-to-noise ratios. Therefore, most studies have used manual preselection of high-quality sounds to achieve better classification rates. Consequently, most systems have never been validated under realistic field conditions. 2. In this study, we evaluated the performance of a passive acoustic monitoring system for four primate species in the highly noisy rain forest environment of the Taï National Park, Côte d'Ivoire. We collected 12 851 h of recordings with 20 autonomous recording units and did not preselect high-quality sounds manually. To automatically detect and classify the sounds of interest, we used an automated system built on speaker segmentation, support vector machines and Gaussian mixture models. One hundred and seventy-nine hours of recordings were used for validating the system. 3. The system performed well in detecting the loud calls of Cercopithecus diana and Colobus polykomos with a recall of 50% and 42%, respectively. Recall rates were lower for Pan troglodytes and Procolobus badius. To determine the presence of Cercopithecus diana and Colobus polykomos with a certainty of P > 0·999, 2 and 7 h of recordings were needed, respectively. For these two species, our automated approach reflected the spatio-temporal distribution of vocalization events well. Despite the seemingly low precision, time investment for the manual removal of false positives in the system's output was only 3·5% compared to a human collecting and processing the primate vocalization data. 4. The proposed monitoring system is already fully applicable for Cercopithecus diana and Colobus polykomos, whereas it needs further improvement for the other species tested. In principle, it can be applied to any distinctive animal sound and can be implemented for the collection of acoustic data for behavioural, ecological and conservation studies.