We propose a new method for analysing the time-frequency domain, called LiFT. It is especially designed for partial tracking in a polyphonic automatic transcription model. After the signal passes through a Q-constant filter bank, composed of twenty four quartertone filters, it is analysed thanks to a generalized maximum likelihood approach. Two hypotheses are tested: the first one is that the output signal of a filter is a cosine plus noise, the second one is that it corresponds to colored noise. This likelihood analysis is developed in two ways: temporally treating the samples and frequentially treating the short time Fourier transform of the signal. For these two approaches, we have tested the robustness to noise and the cosine detection power.

Download