COMPREHENSIVE EVALUATION OF SINGING QUALITY

Abstract

Singing is a popular medium of entertainment, and a desirable skill to develop. But singing pedagogy remains heavily dependent on human music experts. In this thesis, we study the methodology for automatic evaluation of singing voice with respect to two broad aspects of singing quality - prosody and pronunciation.

We study various prosody-related parameters of singing quality judgment, as identified by music experts, such as pitch, rhythm, vibrato, and timbre, and objectively characterize them. We incorporate a cognitive modeling theory inspired by the telecommunication standard PESQ (Perceptual Evaluation of Speech Quality) to provide a perceptually valid objective score to assess singing quality. Furthermore, we design a reference-independent method of ranking singers that leverages on the large amounts of singing data available to compute inter-singer statistics.

Another aspect of singing quality is pronunciation. We address the problem of the lack of lyrics-aligned singing vocals datasets by designing a strategy to automatically build such datasets with the help of imperfect transcriptions from automatic speech recognition (ASR) and non-aligned published lyrics. With this dataset, we build singing-adapted speech acoustic models and show their usability in pronunciation evaluation in singing voice and automatic lyrics-to-audio alignment.