We previously proposed the Auto-Regressive Hidden Markov Model (AR-HMM) for speech signal analysis, where the HMM was introduced as a non-stationary glottal source model. In this paper, we propose a novel method that can automatically generate the topology of the Glottal Source Hidden Markov Model (GS-HMM), as well as estimate the AR-HMM parameter obtained by combining the AR-HMM parameter estimation method and the Minimum Description Length-based Successive State Splitting (MDL-SSS) algorithm. In the experiments, we apply the proposed method to analyze the laryngeal and esophageal voices. The topology generated from the laryngeal voices tended to form a ring state, compared with the topology of the esophageal voices; this result indicates that the glottal sources of the laryngeal voices exhibit clearer periodicity than the sound sources of the esophageal voices. We also compared the vocal tract characteristics estimated by the proposed method and a conventional LP method. From these results, we were able to confirm the feasibility and the validity of the proposed method.
Index Terms: glottal source, AR-HMM, MDL-SSS
Bibliographic reference. Sasou, Akira (2012): "Automatic topology generation of glottal source HMM", In INTERSPEECH-2012, 1616-1619.