A REAL-TIME BEAT TRACKER FOR UNRESTRICTED AUDIO SIGNALS

Analysis of real music content, not available in symbolic form, still remains a very challenging problem. Promising results can be obtained combining signal processing techniques with intelligent agents, in order to support the often ambiguous results of the analytic phase with smart decision systems, trained by a consistent preliminary knowledge or characterized by forms of learning. In this paper we propose a multi- agent algorithm for beat and tempo analysis and induction for unrestricted audio signals; it is based on the combination of lossy onset detection, note accentuation evaluation to estimate metrically essential events, and a multi-agent mechanism to allow dynamic beat tracking. Each agent maintains a self-confidence attribute to rate the confidence for the theory it supports. Consistent test criteria have been used. Experimental results are reported for a database of musical samples from different styles and genres; these results are quite promising. The integration with a harmony analyzer for mutual consolidation is envisaged as next step.