Estimating Japanese word accent from syllable sequence using support vector machine

This paper proposes two methods that estimate, from the word reading (syllable sequence), the place in the word where the accent should be placed (hereafter we call it “accent type”). Both methods use a statistical classifier; one directly estimates accent type, and the other first estimates tone high and low labels and then decides the accent type from the tone label sequence obtained before. Experiments show that both offer high accuracy in the estimation of accent type of Japanese proper names without the use of linguistic knowledge.