Acoustic Signal Extraction and Enhancement

点击翻译按钮获取中文摘要

This project is dedicated to advancing intelligent algorithms for signal extraction and enhancement using acoustic sensor networks (ASNs). Representing Layer 2 in the three-layered approach followed by the proposed research unit, it is based on the ASN infrastructure provided by Layer 1, and delivers enhanced signals of possibly multiple target sources of interest to Layer 3 to allow acoustic scene analysis and understanding. As a core component serving the entire research unit, this project establishes and augments a so-called acoustic map which represents the current state and the dynamics of the acoustic scenario. As basic information, this acoustic map includes attributes of sensors, sources and the acoustic environment, from which higher-level information is derived, e.g., the utility of a given sensor node for a specific task related to a certain source. The entries in the acoustic map are a result of parameter estimation, signal processing, and data-driven learning algorithms and also incorporate prior knowledge. They will be represented by a probabilistic framework and thus include reliability information, which is then heavily exploited for the tasks of signal extraction and enhancement in this project and all other projects. Beyond the blind and semi-blind spatiotemporal filtering algorithms of the preceding project phase, the use of additional reference information for signal enhancement is emphasized in the proposed project. First, acoustic echo cancellation (AEC) will be generalized to the multiple-input/multiple-output (MIMO) case as given by multiple loudspeakers and multiple microphones in the ASN scenario. Here, the acoustic map information will determine for which of the individual loudspeaker-enclosure-microphone paths the classical AEC paradigm of supervised multichannel adaptive filtering is applicable and useful. For other sources of interference, reference information, such as location, activity patterns or spatial and spatiotemporal features, will be estimated or learned by the most useful sensors and shared via the acoustic map so that it can be optimally used throughout the ASN. Adding to the potential of distributed sensing, mobile sensor nodes, e.g., attached to robots, allow exploration and improved coverage of dynamic scenarios. The resulting time-varying sensor array topologies will be optimized for source localization and signal enhancement by selecting optimum subsets of fixed sensors and by optimizing a robot’s trajectories. Accounting for the distributed sensing and distributed processing capacity of the network infrastructure provided by P1, distributed algorithms for the acquisition and fusion of acoustic map information and for efficient AEC and signal enhancement will be designed. To verify their performance in realistic scenarios, selected novel distributed algorithms will be incorporated into a real-time demonstrator that will be jointly developed by all projects.