Detection of Prosodics by Using a Speech Recognition System

摘要

The problem was to determine the ability of a speech recognizer to extractprosodic speech features, such as pitch and stress, and to examine these features for application to future voice recognition systems. The Speech Systems Incorporated (SSI) speech recognizer demonstrated that it could detect prosodic features and that these features do indicate the word and/or syllable that is stressed by the speaker. The research examined the effect of prosodics, such as pitch, amplitude, and duration, on word and syllable stress by using the SSI. Subjects read phases and sentences, using a given intonation and stress. The three sections of the experiment compared questions and answers, words stressed within a sentence, and noun/verb pairs, such as object and subject. The results were analyzed both on the syllable level and the word level. In all cases, there was a significant increase in pitch, amplitude, and duration when comparing stressed words and syllables to unstressed words and syllables. When comparing unstressed words only, it was also noted that the first word in a sentence has an increase in pitch, amplitude, and duration. The threshold could be set in recognition systems for each of these parameters. Current speech recognizers do not use acoustic data above the word level. This research shows that we have the capability of developing better speech systems by incorporating prosodics with new linguistic software.