Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Connected Spoken Word Recognition Using a Many-State Markov Model

Tomio Takara, Naoto Matayoshi, Kazuya Higa

Department of Information Eng., College of Eng., University of the Ryukyus, Okinawa-ken, Japan

This paper is a report on an application of the Markov model to an automatic speech recognition system, in which a large number of states are adopted to model the transitional characteristics of speech more accurately. Unlike the traditional HMM, the feature vectors of this model are regarded to be the states of the Markov model. The transition-probability of the state is, in its initial condition, assumed to be represented by multidimensional normal density function of the feature vector. The many-state model is obtained by quantizing the feature vector (state) space and sampling the probability density function at each code vector. The resulting recognizer was tested and compared on a vocabulary of four-digit numerals using 3 dimensional feature vector sequences. The many-state model attained a recognition score of 98.2%, which was 1.6% higher than that of a five-state traditional HMM.

Full Paper

Bibliographic reference.  Takara, Tomio / Matayoshi, Naoto / Higa, Kazuya (1994): "Connected spoken word recognition using a many-state Markov model", In ICSLP-1994, 235-238.