Through experiment, this research introduces and verifies the usefulness of prosody attributes such as loudness, pitch and sound length as an emotional feature to express the characteristics of the emotion being felt. Sound length is proportional to pronunciation duration and is inversely proportional to the number of phonemic changes per unit of time. Based on these facts, speech speed and the emotional feature were calculated as follows. First, a codebook was generated using mel-frequency cepstral frequency (MFCC) vectors from a preexisting emotional speech database. Second, the MFCC vector of the speech signal was vector-quantized by this codebook to generate a quantized sequence. Third, this sequence was considered for a phoneme sequence and the speech speed was computed by normalizing the number of phoneme changes for each window of time. Fourth, the emotional feature was generated based on this speech speed as follows. The speech speed was added to the MFCC vector with delta and acceleration computation to generate an emotional feature that is related to prosody elements such loudness, pitch and sound length. In order to analyze the utility of these emotional features, a recognition system was developed with the emotional features and a hidden Markov model (HMM). For maximum performance, the degree of MFCC, size of the codebook, method of speech speed computation, window size of speech speed computation, number of HMM model state and the number of the Gaussian mixture model (GMM) per state were carefully selected. To test the recognition system a text-independent, speaker-independent experiment and a text-independent, speaker-dependent experiment were conducted. It was verified that the recognition system using emotional features showed better performance than the recognition system using only speech features, with improvements of 2.5% and 3.5%, respectively, in different experiments.

http://jcse.kiise.org/PublishedPaper/topic_abstract.asp?idx=341