Novel time-frequency representations for music information retrieval

Abstract

Time-frequency representations (TFR) are one of the most valuable tools in digital audio processing, being used in many applications. TFRs can be computed having different time or frequency resolutions and can even represent a certain frequency variation over time, e.g. when using the fan-chirp transform. The main shortcoming of TFRs is the energy smearing related to non-stationarity of the signal within the analysis windows used. This kind of artifact usually results in performance degradation of applications that make use of TFRs, hence providing TFRs that precisely represent the signals of interest is crucial to enhance the performance of such systems. A way to compute a high-resolution TFR is to combine TFRs having different resolutions in such a way that preserves the best aspects of each representation. This is the general idea behind all methods proposed in this thesis, of which the main goal is to allow for a sharp representation of main melody signals in polyphonic contexts. The methods are classified as: bin-wise combinations, combinations based on local information, and methods based on image analysis. Their performance are assessed by means of several experiments using both synthetic and real-world signals, and the results indicate the proposed multi-resolution fan-chirp interpolation method as the best in terms of frequency bandwidth, onset definition and dynamic range. Also, an automatic annotation scheme was devised to diminish the human effort in the transcription of rhythm patterns. This method utilizes TFRs with coarse frequency resolution and a clustering procedure to classify the types of hit. The estimated accuracy in terms of classification is around 75% to 80%.