Mirex 2012 Audio Beat Tracking Submission: Ibt

This extended abstract briefly describes our submission for the “Audio Beat Tracking” task. The proposed system is fully described in [Oliveira et al., “Beat tracking for multiple applications: a multi-agent system architecture with state recovery.” IEEE Transactions on Audio Speech and Language Processing, 20(10):1-10, in press, 2012]. The proposed system integrates an automatic monitoring and state recovery mechanism, that applies (re-)inductions of tempo and beats, on a multi-agent-based beat tracking architecture. Beats can be predicted in a causal or in a noncausal usage mode, which makes the system suitable for diverse applications. 1. SYSTEM DESCRIPTION IBT (standing for INESC Porto Beat Tracker) is the proposed tempo induction and beat tracking algorithm, fully described in [1]. It is inspired by the multi-agent tracking architecture of BeatRoot, where competing agents process parallel hypotheses of tempo and beat [2]. As depicted in Fig. 1, IBT’s algorithm follows a topdown architecture composed of: i) an audio feature extraction module that parses the audio data into a continuous feature sequence assumed to convey the predominant information relevant to rhythmic analysis; followed by ii) an agents induction module, which (re-)generates a set of new hypotheses regarding possible beat periods and phases; followed by iii) a beat tracking module, which propagates hypotheses, proceeds to their online creation, killing and ranking, and outputs beats on-the-fly and/or at the end of the analysis. To handle abrupt changes in the musical signal more rapidly and robustly, in real-time contexts (e.g., data streaming), the system also integrates iv) an automatic monitoring mechanism (AMM). This mechanism is responsible for supervising the beat tracking analysis of the signal to the necessity of recovering the state of the system through re-inductions of beat and tempo. This document is licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 License. http://creativecommons.org/licenses/by-nc-sa/3.0/ c © 2012 The Authors. Figure 1. IBT block diagram. 1.1 Audio Feature Extraction Our implementation makes use of the spectral flux as our mid-level rhythmic representation. The spectral flux measures magnitude variations across all frequency bins, k, of the signal’s spectrum,X(n, k), along consecutive analysis frames, n. We compute the time-frequency representation of the signal through a Fast Fourier Transform (FFT), using a Hamming window envelope with w = 1024 samples (23.2ms at a sampling rate of Fs = 44100Hz) and 50% overlap. As proposed in [3], the spectral flux is calculated using the L1-norm over a linear magnitude, which is halfwave rectified, HWR(x) = x+|x| 2 , to retain only increasing variations in the magnitude spectrum: