A Change Discrimination Onset Detector with Peak Scoring Peak Picker and Time Domain Correction

An onset detector is described based on the most successful onset detector for non-pitched percussive audio events from an earlier comparative study (Collins, 2005a). The detection function is an adaptation of the log intensity difference change discrimination originally introduced by Anssi Klapuri (Klapuri, 1999). A novel peak picking method is used based on scoring the most salient peaks with respect to the local function terrain. Discovered onset positions are corrected using parallel finer resolution time domain methods. The implementation is much faster than realtime and causal, being a conversion of code released for a real-time computer music system. In the MIREX evaluation, across both percussive and non-percussive onsets, the algorithm performed well on the former but was not successful on the latter. It was the most efficient and had the best overall time localisation of onsets of the algorithms in the contest. 1 Algorithm Overview A real-time causal onset detector had been developed for live computer music use, using the most successful onset detector for non-pitched percussive (NPP) samples from a comparative study (Collins, 2005a). This algorithm was adapted as an entry for the MIREX 2005 Audio Onset Detection contest. Figure 1 gives an overview of the processing steps in the algorithm. 2 Detection Function The detection function is an adaptation of the log intensity difference change discrimination originally introduced by Anssi Klapuri (Klapuri, 1999). Processing is FFT based and involves an intensity to decibel transform in ERB scale bands with equal loudness contour correction and a general averaged difference over the last three frames (Collins, 2005a). The onset detector calculates a 1024 point FFT with hop size of 512, assuming target 44100Hz audio.