Automatic alignment of long syllables in a cappella Beijing opera

Abstract

In this study we propose how to modify a standard approach for text-to-speech alignment to apply in the case of alignment of lyrics and singing voice. We model phoneme durations by means of a duration-explicit hidden Markov model (DHMM) phonetic recognizer based on MFCCs. The phoneme durations are empirically set in a probabilistic way, based on prior knowledge about the lyrics structure and metric principles, specific for the Beijing opera music tradition. Phoneme models are GMMs trained directly on a small corpus of annotated singing voice. The alignment is evaluated on a cappella material from Beijing opera, which is characterized by its particularly long syllable durations. Results show that the incorporation of music-specific knowledge results in a very high alignment accuracy, outperforming significantly a baseline HMM-based approach.