Temporal parameters discriminate better between read and narrated speech in Brazilian Portuguese

This work shows that 5 out of 6 acoustic parameters that correctly classify read and narrated speech in Brazilian Portuguese are temporal parameters. Several statistical models showed that significant differences between the styles are revealed by: speech rate, a measure of articulation rate, duration-related salience rate, mean and standard-deviation of degree of duration-related salience and mean of F0 first derivative. A set of 161 excerpts of narrated and read speech from ten speakers was used for training an LDA model. Another set of 57 excerpts with different subjects was used for testing the same model. Its performance with the six aforementioned parameters has achieved in the case of read speech an accuracy rate of 90 % for the training subset and 94 % for the test subset and in the case of narrated speech 70 % for the training subset and 27 % for the test subset.