Location Of Phrase Boundaries Research Articles

Perceived prosody: Phonetic bases of prominence and boundaries.

In comprehending speech, listeners are sensitive to the prosodic features that signal the phrasing and the discourse salience of words (prominence). Findings from two experiments on prosody perception show that acoustic and articulatory kinematic properties of speech correlate with native listeners’ perception of phrasing and prominence. Subjects in this study were 114 university-age adults (74 UIUC + 40 Haskins), monolingual speakers of American English who were untrained in prosody transcription. Subjects listened to short recorded excerpts (about 20 s) from two corpora of spontaneous and read speech (Buckeye Corpus and Wisconsin Microbeam Database) and marked prominent words and the location of phrase boundaries on a transcript. Intertranscriber agreement rates across subsets of 17–40 subjects are significantly above chance based on Fleiss’ statistic, indicating that listeners’ perception of prosody is reliable, with higher agreement rates for boundary perception than for prominence. Prosody perception varies across listeners (both corpora) and across speakers (WMD, where perceived prosody varies for the same utterance produced by different speakers). Acoustic measures from stressed vowels (Buckeye: duration, intensity, F1, F2) and articulatory kinematic measures (WMD) are correlated with the perceived prosodic features of the word. [Work supported by NSF.]

Automatic ToBI prediction and alignment to speed manual labeling of prosody

Tagging of corpora for useful linguistic categories can be a time-consuming process, especially with linguistic categories for which annotation standards are relatively new, such as discourse segment boundaries or the intonational events marked in the Tones and Break Indices (ToBI) system for American English. A ToBI prosodic labeling of speech typically takes even experienced labelers from 100 to 200 times real time. An experiment was conducted to determine (1) whether manual correction of automatically assigned ToBI labels would speed labeling, and (2) whether default labels introduced any bias in label assignment. A large speech corpus of one female speaker reading several types of texts was automatically assigned default labels. Default accent placement and phrase boundary location were predicted from text using machine learning techniques. The most common ToBI labels were assigned to these locations for default tones and break type. Predicted pitch accents were automatically aligned to the mid-point of the word, while breaks and edge tones were aligned to the end of the phrase-final word. The corpus was then labeled by a group of five trained transcribers working over a period of nine months. Half of each set of recordings was labeled in the standard fashion without default labels, and the other half was presented with preassigned default labels for labelers to correct. Results indicate that labeling from defaults was generally faster than standard labeling, and that defaults had relatively little impact on label assignment.

Spectral Distribution of Prosodic Information

Prosodic speech cues for rhythm, stress, and intonation are related primarily to variations in intensity, duration, and fundamental frequency. Because these cues make use of temporal properties of the speech waveform they are likely to be represented broadly across the speech spectrum. In order to determine the relative importance of different frequency regions for the recognition of prosodic cues, identification of four prosodic features, syllable number, syllabic stress, sentence intonation, and phrase boundary location, was evaluated under six filter conditions spanning the range from 200-6100 Hz. Each filter condition had equal articulation index (AI) weights, AI = 0.01; p(C)isolated words approximately equal to 0.40. Results obtained with normally hearing subjects showed that there was an interaction between filter condition and the identification of specific prosodic features. For example, information from high-frequency regions of speech was particularly useful in the identification of syllable number and stress, whereas information from low-frequency regions was helpful in identifying intonation patterns. In spite of these spectral differences, overall listeners performed remarkably well in identifying prosodic patterns, although individual differences were apparent. For some subjects, equivalent levels of performance across the six filter conditions were achieved. These results are discussed in relation to auditory and auditory-visual speech recognition.

The transmission of prosodic information via selected spectral regions of speech

In a recent article [K. W. Grant and L. D. Braida, J. Acoust. Soc. Am. 89, 2952–2960 (1991)] it was demonstrated that spectrally different bands of speech with equal articulation index (AI) scores provided approximately equal auditory-visual sentence recognition when combined with speechreading. Given that different parts of the frequency spectrum provide different segmental cues for consonant and vowel recognition, current models of auditory-visual integration [e.g., L. D. Braida, J. Exp. Psychol. 43A, 647–677 (1991)] would predict that some spectral regions of speech are more complementary to speechreading than others. This raises the possibility that the findings of Grant and Braida are attributable, at least in part, to suprasegmental cues that were transmitted equally well by the different spectral bands tested. To test this possibility, the identification of syllable number, syllable stress, sentence intonation, and phrase boundary location was assessed under six approximately equal AI filter conditions similar to those evaluated by Grant and Braida. The results indicate that syllable number and syllable stress are perceived best through high-frequency bands, intonation is perceived best through low-frequency bands, and phrase boundary location is perceived equally well throughout the speech spectrum. These results are discussed in terms of the importance of different spectral regions of speech for the recognition of suprasegmental cues, and how this may relate to overall speech intelligibility. [Work supported by NIH Grant No. DC00792.]

Improvement of synthetic speech quality through syntactic information

Many words in Japanese have identical written expression but different pronunciation. Natural synthetic speech therefore requires selection of the correct pronunciation for words and optimized prosodic features, including accent position and level, sentence intonation and length of pause, through the use of syntactic features. This paper describes (1) a new method of determining phrase accent level, based on accentual phrase boundary location and compound word structure, and (2) a newly proposed syntactic class of phrase boundaries. The results of the automatic determination of pronunciation, and opinion tests of intelligibility and naturalness are also described. About 10 000 words are assigned to syntactic and semantic features to determine correct pronunciation, representing about 20% of the total vocabulary. Pronunciation of 99% of the words in a Japanese economic daily were correct, and naturalness of the synthetic speech was 1.1 grades higher under the five-grade opinion test.