Analysis of expression in singing voice is gaining more importance as the current assessment systems fail to consider important resources in expressive singing, e.g. phonation modes. Phonation modes have been divided into four categories (breathy, pressed, neutral and flow) that correspond to levels of glottal adduction force. This thesis focuses on the analysis and automatic classification of phonation modes, and proposes a visual feedback system designed for singing voice assessment, vocal education and musicological analysis. We propose to use a wide range of audio descriptors in order to extract information from the audio signal and to perform feature selection for reducing the dimension of the feature set. A supervised classification approach is applied with making use of Multi-Layer Perceptrons (MLP). The hyperparameters of the model are optimized with cross validation on training subsets. The results of the evaluation of the obtained model outperform the state of the art methods. In order to generalize the feature analysis to avoid bias caused by having insufficient data we curated two new datasets for phonation modes research. Finally, the designed visual feedback system is tested with singing students and teachers to assess its usefulness for educational purposes.