Automatic transcription of bass guitar tracks applied for music genre classification and sound synthesis

Music recordings most often consist of multiple instrument signals, which overlap in time and frequency. In the field of Music Information Retrieval (MIR), existing algorithms for the automatic transcription and analysis of music recordings aim to extract semantic information from mixed audio signals. In the last years, it was frequently observed that the algorithm performance is limited due to the signal interference and the resulting loss of information. One common approach to solve this problem is to first apply source separation algorithms to isolate the present musical instrument signals before analyzing them individually. The performance of source separation algorithms strongly depends on the number of instruments as well as on the amount of spectral overlap. In this thesis, isolated instrumental tracks are analyzed in order to circumvent the challenges of source separation. Instead, the focus is on the development of instrument-centered signal processing algorithms for music transcription, musical analysis, as well as sound synthesis. The electric bass guitar is chosen as an example instrument. Its sound production principles are closely investigated and considered in the algorithmic design. In the first part of this thesis, an automatic music transcription algorithm for electric bass guitar recordings will be presented. The audio signal is interpreted as a sequence of sound events, which are described by various parameters. In addition to the conventionally used score-level parameters note onset, duration, loudness, and pitch, instrument-specific parameters such as the applied instrument playing techniques and the geometric position on the instrument fretboard will be extracted. Different evaluation experiments confirmed that the proposed transcription algorithm outperformed three state-of-the-art bass transcription algorithms for the transcription of realistic bass guitar recordings. The estimation of the instrument-level parameters works with high accuracy, in particular for isolated note samples. In the second part of the thesis, it will be investigated, whether the sole analysis of the bassline of a music piece allows to automatically classify its music genre. Different score-based audio features will be proposed that allow to quantify tonal, rhythmic, and structural properties of basslines. Based on a novel data set of 520 bassline transcriptions from 13 different music genres, three approaches for music genre classification were compared. A rule-based classification system could achieve a mean class accuracy of 64.8 % by only taking features into account that were extracted from the bassline of a music piece. The re-synthesis of a bass guitar recordings using the previously extracted note parameters will be studied in the third part of this thesis. Based on the physical modeling of string instruments, a novel sound synthesis algorithm tailored to the electric bass guitar will be presented. The algorithm mimics different aspects of the instrument’s sound production mechanism such as string excitement, string damping, string-fret collision, and the influence of the electro-magnetic pickup. Furthermore, a parametric audio coding approach will be discussed that allows to encode and transmit bass guitar tracks with a significantly smaller bit rate than conventional audio