8.3.1 Prosodic Structure

Spoken sentences have prosodic structure in the sense that some words seem to group naturally together and some words seem to have a noticeable break or disjuncture between them. Often prosodic structure is described in terms of prosodic phrasing, meaning that an utterance has a prosodic phrase structure in a similar way to it having a syntactic phrase structure. For example, in the sentence I wanted to go to London, but could only get tickets for France there seems to be two main intonation phrases, their boundary occurring at the comma. Furthermore, in the first phrase, there seems to be another set of lesser prosodic phrase boundaries (often called intermediate phrases) that split up the words as follows I wanted to go to London.

Prosodic phrasing has many implications for speech synthesis; the final vowel of a

phrase is longer than usual, we often insert a pause after an intonation phrases, and, as we will discuss in Sec. 8.3.6, there is often a slight drop in F0 from the beginning of an intonation phrase to its end, which resets at the beginning of a new intonation phrase.

Practical phrase boundary prediction is generally treated as a binary classification task, where we are given a word and we have to decide whether or not to put a prosodic boundary after it. A simple model for boundary prediction can be based on deterministic rules. A very high-precision rule is the one we saw for sentence segmentation: insert a boundary after punctuation. Another commonly used rule inserts a phrase boundary before a function word following a content word.

More sophisticated models are based on machine learning classifiers. To create a training set for classifiers, we first choose a corpus, and then mark every prosodic boundaries in the corpus. One way to do this is prosodic boundary labeling is to use an intonational model like ToBI or Tilt (see Sec. 8.3.4), have human labelers listen to speech and label the transcript with the boundary events defined by the theory. Because prosodic labeling is extremely time-consuming, however, a text-only alternative is often used. In this method, a human labeler looks only at the text of the training corpus, ignoring the speech. The labeler marks any juncture between words where they feel a prosodic boundary might legitimately occur if the utterance were spoken.

Given a labeled training corpus, we can train a decision tree or other classifier to make a binary (boundary vs. no boundary) decision at every juncture between words (Wang and Hirschberg, 1992; Ostendorf and Veilleux, 1994; Taylor and Black, 1998).

Features that are commonly used in classification include:

• Length features: phrases tend to be of roughly equal length, and so we can use various feature that hint at phrase length (Bachenko and Fitzpatrick, 1990; Grosjean et al., 1979; Gee and Grosjean, 1983).

– The total number of words and syllables in utterance

• The distance of the juncture from the beginning and end of the sentence (in words or syllables)

– The distance in words from the last punctuation mark

• Neighboring part-of-speech and punctuation:

The part-of-speech tags for a window of words around the juncture. Generally the two words before and after the juncture are used.

– The type of following punctuation

There is also a correlation between prosodic structure and the syntactic structure that will be introduced in Ch. 12, Ch. 13, and Ch. 14 (Price et al., 1991). Thus robust parsers like Collins (1997) can be used to label the sentence with rough syntactic information, from which we can extract syntactic features such as the size of the biggest syntactic phrase that ends with this word (Ostendorf and Veilleux, 1994; Koehn et al., 2000).