This research is concerned with machine perception and analysis of complex sound signals. The goaL is to reliably identify and track simultaneous acoustic sources in a monaural signal. In the proposed architecture, perception results from the interaction of data-driven and expectation-driven agents. The allocation of resources to systems agents, and the control of feedback loops between different levels of interpretation of the time-varying signal are approached with strategies that simulate real-time problem solving. At the lowest level of analysis, multi-rate signal processing, used in conjection with focus-switching heuristics, ought to yield high resolution simultaneously in time and frequency, thus giving an efficient method for improving upon traditional bandwidth-time tradeoffs. Source coherence criteria derived from psychoacoustic work on auditory streaming (including correlated AM and FM modulation among partials) are expected to be useful for separating sources when more familiar methods do not suffice. The proposed system relies on a learning co-processor to attune itself to increasingly elusive aspects of a signal. Relevant techniques include traditional parameter adaptation, numerical taxonomy, syntatic pattern matching and concept learning. Especially important is the development of hybrid methods that combine parametric and structural views. In summary, the research addresses key areas of acoustic analysis as well as broader issues in the architecture of artificial intelligence.
