Typical biological environments comprise complex mixtures of signals from multiple biological and environmental sources. Some sources contain critical information that researchers seek to acquire; other sources are distractions that interfer with data acquisition. Common acoustic environments are an important example: a significant segment of the aging US population has difficulty coping with noisy settings. Current solutions are limited to hearing aids, which notoriously amplify all sounds, or headsets, which selectively amplify a single source but isolate the listener from the rest of his or her acoustic environment. Our goal is to enable the development of health-related applications by providing a software library and a turn-key instrument that enable biomedical researchers to easily isolate information-bearing signals from interfering maskers. We propose to develop a system called DMX that uses innovative signal processing techniques to isolate and extract (that is, demix) individual acoustic and bioelectrical source signals from the output of multiple sensors that are generally responding with an unknown mixture of simultaneous sources. The core DMX algorithm we have implemented, and whose effectiveness we demonstrated in Phase 1, cleans up live signals in real time by separating competing foreground sources, and suppressing background noise. A proven DMX innovation is the use of taggers. A tagger is a sensor attached to a significant target or masking source that is identified to the system. Other sensors detect remote (untagged) targets or noise sources. Our current DMX algorithm constitutes a general-purpose blind source separation (BSS) algorithm that advances the state of the art. In Phase 2, we propose to package this algorithm as a fully tested, documented, supported, and deployable software library with MATLAB, C++, and Python interfaces. The library will provide reliable BSS capability to the research community, as well as to designers of assistive listening devices. The library will also be suitable for processing bioelectric signals - EEG, EMG, etc. - to allow researchers and clinicians to isolate sources of interest from response mixtures (e.g. fetal and maternal heartbeats). We will also develop and sell a turn-key DMX instrument, complete with up to eight microphones, signal processing electronics, and control software. This version of DMX will be useful to researchers who need to produce high-fidelity low-noise recordings in noisy environments such as MRI scanners, and who are not audio or bioelectrical signal engineers. This instrument will allow such a user to tag the most prominent sources, record the entire signal scene, and extract the separate source signals and related location information. In Phase 2 we aim to reduce source separation time by employing dynamic error analysis, the intelligent use of environmental information such as source-to-sensor distance information, and the reuse of previously generated separation solutions. Both versions of the DMX product will be ready for commercial use by the end of the project.
This project develops DMX, a product to promote the progress of biomedical research by providing well- tested reliable signal processing software and instruments for extracting 'pure' reconstructions of information- bearing signals in noisy biomedical environments. Such signals of interest are usually mixed with noise and distorted by reflections, and often masked by interfering signals. DMX uses innovative signal processing techniques to disentangle the 'babble' of competing signal mixtures into separate well-defined channels, each of which represents a single isolated signal source in the environment.
