Singing voice recognition considering high-pitched and prolonged sounds

A conventional Large Vocabulary Continuous Speech Recognition (LVCSR) system has difficulty recognizing singing voices accurately because both the high-pitched and prolonged sounds of singing voices tend to degrade its recognition accuracy. We previously described an Auto-Regressive Hidden Markov Model (AR-HMM) and an accompanying parameter estimation method. We demonstrated that the AR-HMM accurately estimated the characteristics of both articulatory systems and excitation signals from high-pitched speech. In this paper, we describe an AR-HMM applied to feature extraction from singing voices and propose a prolonged-sound detection and elimination method.