Automatic Assessment of Singing

TitleAutomatic Assessment of Singing
Publication TypeMaster Thesis
Year of Publication2019
AuthorsPuranik, N. V.
AbstractThis thesis aims to develop an automatic singing evaluation system specially suited to evaluate notes singing exercises. We build Hindustani singing datasets with a combined collection of audio samples of 349 reference and performance pairs. The samples are annotated with an overall grade based on pitch accuracy. For this purpose, we develop a GUI grading tool which provides a visual feedback for the performance. This tool helps reduce human biases in the grade annotations. The existing baseline system (BMCS) for singing assessment developed using Turkish Conservatory dataset (MAST dataset) is extensively studied to identify audio alignment as one of the possible areas of improvement. A methodology and appropriate metrics are devised to test the audio alignment performance. Using this methodology, different features are tested to demonstrate that an improved audio-to-audio alignment system can be achieved using a 120-dimensional HPCP feature. The second part of this study is focused on finding suitable features to assess a singing performance, given a good alignment of reference and student audio. A novel 'pitch-histogram cosine distance' feature is devised to measure note-level accuracy of singing. The effectiveness of these features with respect to the baseline features is shown by linear regression models trained and tested using the Hindustani and MAST datasets. The effectiveness of 'pitch-histogram cosine distance' is indicated by the low mean absolute errors and the interpretability of the linear models developed. These features are also used to provide a note-level accuracy visualization of student performance.
intranet