Abstract
Singing Voice is the oldest musical instrument, the sound of which is uttered from the collaboration of our respiratory system, our larynx and our vocal tract. We perceive the sound as music because we can give a precise, concise and scienti?c measurement on its musical features, such as pitches (e.g. melody and harmony), rhythm (e.g. tempo and articulation) and dynamics (e.g. loudness and softness). Singing voice analysis, espe-cially in popular music, leads to a plentiful potential and commercial applications, such as singing skill evaluation, singing style visualization, singer identi?cation, singing synthesis and singing voice separation. In this thesis, we propose to solve several research problems associated with singing voice analysis in popular music using machine learning approaches. We analyze 4 tasks related to the singing voice, namely Evaluating Singing, Visualizing Singing, Singing Fea-tures Engineering and Singing Voice Separation. Our proposed solutions would further advance the machine to become an android, who can appreciate and manipulate the singing voice in the way which is similar to what we, the human being, do. In the task of Evaluating Singing, we develop a self-pitch-training iOS app to quan-titatively investigate the ef?cacy of the real-time interactive visual feedback and the singing demonstration, along with the user preference of these learning aid features. This app serves as a guideline for implementing and enhancing similar training apps. In the task of Visualizing Singing, we modify a trajectory clustering algorithm called TRACLUS to avoid the problems found in a well-known visualizing technique called performance worm. Further, our technique is able to visualize the singing style under common musical events. The plots produced is able to indicate the distinct stylistic differences between the singers. In the task of Singing Features Engineering, we study the Mel-frequency Cepstral Co-ef?cients (MFCC) and the sinusoidal partials. MFCC is a well-known feature for voice discrimination problem (e.g. words and singers identi?cation) as it characterizes the timbre of vocal voice well. We model the mutual information between the voiceprint and the optimal number of MFCC. The experimental results show that our model is a more comprehensive feature selection criterion for the MFCC than the de-facto cross validation technique. Sinusoidal partials are the deterministic part of an audio sig-nal. They are created from spectral modeling with sinusoidal partials tracking (PT) algorithm and are used for singing synthesis. We propose a simple heuristic-based PT algorithm to track the sinusoidal partials of singing voice in the popular music. The experimental results show that our PT algorithm outperform 2 common benchmarks. Thus, our PT algorithm can be considered as a new benchmark of the PT algorithm used in singing voice analysis. In the task of Singing Voice Separation, we present an image-processing-inspired Convolutional Neural Network (CNN), which is specially designed for separating the singing voice from its musical accompaniment. We show that the proposed CNN out-performs the ?rst runner up in the Music Information Retrieval Evaluation eXchange (MIREX) 2016 and the winner of MIREX 2014 on the iKala dataset. The results also indicate that our model can compete with state-of-the-art singing voice systems which use multi-channel modeling, data augmentation and model blending, when tested on the DSD100 dataset for the full-tracks song evaluation.