Mirex-09 “audio Beat Tracking” Task: Ircambeat Submission

This extended abstract details a submission to the Music Information Retrieval Evaluation eXchange (MIREX) 2009 for the “Audio Beat Tracking” task. The system named ircambeat performs time-variable tempo and meter estimation, beat and downbeat marking. Detailed description of the two parts are given in [1] and [2]. We briefly summarized them below. 1. IRCAMBEAT IMPLEMENTATION Ircambeat is a C++ software and library running under Linux, Windows-XP and Mac-OS-X which performs timevariable tempo and meter estimation, beat and downbeat marking. 2. IRCAMBEAT ALGORITHM DESCRIPTION The flowchart of ircambeat is represented in Figure 1. 2.1 Tempo and meter estimation The tempo and meter estimation algorithm works in three stages. First, an onset-energy-function f(t) is extracted from the audio signal by computing a reassigned spectral-energyflux (using time and frequency reassignement for better precision). Second, the dominant periodicities of f(t) over time are estimated using a combination of Discrete Fourier Transform and Frequency-Mapped Auto-Correlation-Function. The combination of both allows to better emphasizing the periodicities due to the meter, the beat and the tatum periodicities in f(t). We note p(t) the resulting function. Finally, a Viterbi decoding algorithm is used to decode simultaneously the tempo and the meter. For this, we define states of a hidden Markov model as all the combinations of possible tempi and meter (among 22: binary grouping of beat/ binary subdivision of beat, 23: binary/ ternary and 32: ternary/ binary). Given p(t), we compute the observation probabilities of the states over time. The decoding then produces the best estimates of tempo and meter over time. More details about the algorithm can be found in [1]. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. c © 2009 International Society for Music Information Retrieval. Figure 1. Flowchart of ircambeat 2.2 Beat and downbeat tracking Beat and downbeat positions are estimated simultaneously using an inverse Viterbi formulation. In this formulation, a state is defined as a specific time. Observation probabilities of states (times) are obtained using a LDA-trained beat-template. This beat-template is obtained by considering the function f(t) inside a measure as a N-dimensional feature vector. A two-class (beat/ non-beat) problem is then solved using LDA and a training set. The resulting LDA-axe is then used as the best beat-template in order to perform discrimination between beat and non-beat positions. More details about the beat estimation algorithm can be found in [2]. Details about the simultaneous estimation of beat and downbeat are not available since the corresponding paper was rejected to ISMIR09. 3. MIREX09 RESULTS AND DISCUSSIONS 3.1 Experiment Two test-sets were used for MIREX-09 evaluation: McKinney Collection: A collection of 160 musical excerpts; the same collection as the one used for the 2006 Audio Tempo Extraction and Beat tasks. Each recording has been annotated by 40 different listeners (39 in a few cases). Sapp’s Mazurka Collection: 322 files drawn from the Mazurka.org dataset put together by Craig Sapp from CHARM / Royal Holloway, University of London. Craig Sapp was also responsible for creating the highquality ground-truth files. Ten performance measures have been used for the evalution: F-measure: the standard calculation as used in onset evaluation but with a ± 70ms window Cemgil: beat accuracy is calculated using a Gaussian error function with 40ms standard deviation Goto: binary decision of correct or incorrect tracking based on statistical properties of a beat error sequence P-score: McKinney’s impulse train cross-correlation method as used in 2006 CMLC: the ratio of the longest continuously correctly tracked section to the length of the file, with beats at the correct metrical level CMLT: the total number of correct beats at the correct metrical level AMLC: the ratio of the longest continuously correctly tracked section to the length of the file, with beats at allowed metrical levels AMLT: the total number of correct beats at allowed metrical levels D (bits): information based criteria based on analysis of a beat error histogram