Linguistic Patterns Detected Through a Prosodic Segmentation in Spontaneous Taiwan Mandarin Speech

This paper proposes that spontaneous speech, segmented into perceptually coherent prosodic constituents, is able to provide plentiful linguistic information in which clear patterns can be observed. We present pioneering studies with empirical and quantitative evidence, supporting the notion that prosodic units can be useful for the automatic processing of spontaneous speech. High inter-labelers’ consistency proves the applicability of human prosodic segmentation. A series of results on spontaneous Taiwan Mandarin speech suggest that linguistic patterns found in different linguistic aspects can in theory be used for processing and understanding spontaneous speech. In an automatic POS tagging experiment, it is demonstrated that transcripts with annotations of prosodic boundaries achieved a slightly better performance than the original transcripts with only the speaker turn annotation. Making use of prosodic boundaries, we can deal with the problem of disfluency more directly. With regard to lexical and discourse cue phrases, we also found them produced frequently and regularly at the prosodic boundaries.