EUROSPEECH 2003 - INTERSPEECH 2003
8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003

        

Automatic Generation of Non-Uniform Context-Dependent HMM Topologies Based on the MDL Criterion

Takatoshi Jitsuhiro (1), Tomoko Matsui (2), Satoshi Nakamura (1)

(1) ATR-SLT, Japan
(2) Institute of Statistical Mathematics, Japan

We propose a new method of automatically creating non-uniform context-dependent HMM topologies by using the Minimum Description Length (MDL) criterion. Phonetic decision tree clustering is widely used, based on the Maximum Likelihood (ML) criterion, and creates only contextual variations. However, it also needs to empirically predetermine control parameters for use as stop criteria, for example, the total number of states. Furthermore, it cannot create topologies with various state lengths automatically. Therefore, we introduce the MDL criterion as split and stop criteria, and use the Successive State Splitting (SSS) algorithm as a method of generating contextual and temporal variations. This proposed method, the MDL-SSS, can automatically create proper topologies without such predetermined parameters. Experimental results show that the MDLSSS can automatically stop splitting and obtain more appropriate HMM topologies than the original one. Furthermore, we investigated the MDL-SSS combined with phonetic decision tree clustering, and this method can automatically obtain the best performance with any heuristic.

Full Paper

Bibliographic reference.  Jitsuhiro, Takatoshi / Matsui, Tomoko / Nakamura, Satoshi (2003): "Automatic generation of non-uniform context-dependent HMM topologies based on the MDL criterion", In EUROSPEECH-2003, 2721-2724.