High-Definition Time-Frequency Representation Based on Adaptive Combination of Fan-Chirp Transforms via Structure Tensor
This paper presents a novel technique for producing high-definition time-frequency representations by combining different instances of short-time fan-chirp transforms. The proposed method uses directional information provided by an image processing technique named structure tensor, applied over a spectrogram of the input signal. This information indicates the best analysis window size and chirp parameter for each time-frequency bin, and feeds a simple interpolation procedure, which produces the final representation. The method allows the proper representation of more than one sound source simultaneously via fan-chirp transforms with different resolutions, and provides a precise reproduction of transient information. Experiments in both synthetic and real audio illustrate the performance of the proposed system.