Isolated word intonation recognition using hidden Markov models

A method is described for recognition of intonation patterns based on discrete distribution hidden Markov models (HMMs) and vector quantization techniques. Fundamental frequency and energy features, were used to determine the best combination of feature processing and quantization techniques for recognition of statement, question, command, calling, and continuation intonation patterns in isolated words. A recognition accuracy of 89% was achieved for the best-case speaker- and word-independent performance. Recognition performance of human listeners on a 100-word subset yielded 77% accuracy, compared to 83% using HMMs on the same subset