8.3.4 More sophisticated models: ToBI

While current synthesis systems generally use simple models of prosody like the ones discussed above, recent research focuses on the development of much more sophisticated

原书第 293 页

cated models. We'll very briefly discuss the ToBI, and Tilt models here.

ToBI

One of the most widely used linguistic models of prosody is the ToBI (Tone and Break Indices) model (Silverman et al., 1992; Beckman and Hirschberg, 1994; Pierrehumbert, 1980; Pitrelli et al., 1994). ToBI is a phonological theory of intonation which models prominence, tune, and boundaries. ToBI's model of prominence and tunes is based on the 5 pitch accents and 4 boundary tones shown in Fig. 8.9.

| --- | --- | --- | --- |

| H $ ^{{*}} $ | peak accent | L-L% | “final fall”: “declarative contour” of American English” |

| L $ ^{{*}} $ | low accent | L-H% | continuation rise |

| L $ ^{{*}} $+H | scooped accent | H-H% | “question rise”: canonical yes-no question contour |

| L+H $ ^{{*}} $ | rising peak accent | H-L% | final level plateau (plateau because H- causes “up-step” of following) |

| H+!H $ ^{{*}} $ | step down | | |

Figure 8.9 The accent and boundary tones labels from the ToBI transcription system for American English intonation (Beckman and Ayers, 1997; Beckman and Hirschberg, 1994).

An utterance in ToBI consists of a sequence of intonational phrases, each of which ends in one of the four boundary tones. The boundary tones are used to represent the utterance final aspects of tune discussed in Sec. 8.3.3. Each word in the utterances can optionally be associated with one of the five types of pitch accents.

Each intonational phrase consists of one or more intermediate phrase. These phrases can also be marked with kinds of boundary tone, including the %H high initial boundary tone, which is used to mark a phrase which is particularly high in the speakers' pitch range, as well as final phrase accents H- and L-.

In addition to accents and boundary tones, ToBI distinguishes four levels of phrasing, which are labeled on a separate break index tier. The largest levels of phrasing are the intonational phrase (break index 4) and the intermediate phrase (break index 3), and were discussed above. Break index 2 is used to mark a disjuncture or pause between words that is smaller than an intermediate phrase, while 1 is used for normal phrase-medial word boundaries.

Fig. 8.10 shows the tone, orthographic, and phrasing tiers of a ToBI transcription, using the praat program. We see the same sentence read with two different intonation patterns. In (a), the word Marianna is spoken with a high H* accent, and the sentence has the declarative boundary tone L-L%. In (b), the word Marianna is spoken with a low L* accent and the yes-no question boundary tone H-H%. One goal of ToBI is to express different meanings to the different type of accents. Thus, for example, the L* accent adds a meaning of surprise to the sentence (i.e., with a connotation like 'Are you really saying it was Marianna?'). (Hirschberg and Pierrehumbert, 1986; Steedman, 2003).

原书第 294 页

Figure 8.10 The same sentence read by Mary Beckman with two different intonation patterns and transcribed in ToBI. (a) shows an H* accent and the typical American English declarative final fall L-L%. (b) shows the L* accent, with the typical American English yes-no question rise H-H%.

ToBI models have been proposed for many languages, such as the J_TOBI system for Japanese (Venditti, 2005); see Jun (2005).

Other Intonation models

The Tilt model (Taylor, 2000) resembles ToBI in using sequences of intonational events like accents and boundary tones. But Tilt does not use ToBI-style discrete phonemic classes for accents. Instead, each event is modeled by continuous parameters that represent the F0 shape of the accent.

Figure 8.11 Schematic view of events in the Tilt model (Taylor, 2000). Each pitch accent (a) and boundary tone (b) is aligned with a syllable nucleus s.

Instead of giving each event a category label, as in ToBI, each Tilt prosodic event is characterized by a set of three acoustic parameters: the duration, the amplitude, and the

原书第 295 页

tilt parameter. These acoustic parameters are trained on a corpus which has been hand-

labeled for pitch accents (a) and boundary tones (b). The human labeling specifies

the syllable which bears the accent or tone; the acoustic parameters are then trained

automatically from the wavefile. Fig. 8.11 shows a sample of a Tilt representation.

Each accent in Tilt is viewed as having a (possibly zero) rise component up to peak,

followed by a (possible zero) fall component. An automatic accent detector finds

the start, peak, and end point of each accent in the wavefile, which determines the

duration and amplitude of the rise and fall components. The tilt parameter is an abstract

description of the F0 slope of an event, calculated by comparing the relative sizes of

the rise and fall for an event. A tilt value of 1.0 indicates a rise, tilt of -1.0 a fall, 0

equal rise and fall, -0.5 is an accent with a rise and a larger fall, and so on:

$$ \begin{aligned}tilt&=\frac{tilt_{amp}+tilt_{dur}}{2}\\&=\frac{|A_{rise}|-|A_{fall}|}{|A_{rise}|+|A_{fall}|}+\frac{D_{rise}-D_{fall}}{D_{rise}+D_{fall}}\end{aligned} $$

See the end of the chapter for pointers to a wide variety of other intonational models.