8.3.6 Computing F0 from prosodic labels

For diphone, articulatory, HMM, and formant synthesis we also need to specify the F0 values of each segment. For the tone sequence models like ToBI or Tilt, this F0 generation can be done by specifying F0 target points for each pitch accent and boundary tone; the F0 contour for the whole sentence can be created by interpolating among these targets (Anderson et al., 1984).

In order to specify a target point we need to describe what it is (the F0 value) and when it occurs (the exact time at which this peak or trough occurs in the syllable). The F0 values of the target points are generally not specified in absolute terms of Hertz. Instead, they are defined relative to pitch range. A speaker's pitch range is the range between the lowest frequency they use in a particular utterance (the baseline frequency) and the highest frequency in the utterance (the topline). In some models, target points are specified relative to a line in between called the reference line.

For example, we might write a rule specifying that the very beginning of an utterance have a target point of 50% (halfway between the baseline and topline). In the rule-based system of Jilka et al. (1999) the target point for an H* accent is at 100% (the topline) and for an L* accent at 0% (at the baseline). L+H* accents have two target points, at 20% and 100%. Final boundary tones H-H% and L-L% are extra-high and extra-low at 120% and -20% respectively.

Second, we must also specify exactly where in the accented syllable the targets apply; this is known as accent alignment. In the rule-based system of Jilka et al. (1999), again, H* accents are aligned 60% of the way through the voiced part of the accent syllable (although IP-initial accents are aligned somewhat later in the syllable, while IP-final accents are aligned somewhat earlier).

Instead of writing these rules by hand, the mapping from pitch accent sequence to F0 value may be learned automatically. For example Black and Hunt (1996) used

原书第 297 页

linear regression to assign target values to each syllable. For each syllable with a pitch accent or boundary tone, they predicted three target values, at the beginning, middle, and end of the syllable. They trained three separate linear regression models, one for each of the three positions in the syllable. Features included:

accent type on the current syllable as well as two previous and two following syllables

lexical stress of this syllable and surrounding syllables

number of syllables to start of phrase and to end of phrase

number of accented syllables to end of phrase

Such machine learning models require a training set that is labeled for accent; a number of such prosodically-labeled corpora exist, although it is not clear how well these models generalize to unseen corpora.

Finally, F0 computation models must model the fact that pitch tends to decline through a sentence; this subtle drop in pitch across an utterance is called declination; an example is shown in Fig. 8.12.

Figure 8.12 F0 declination in the sentence 'I was pretty goofy for about twenty-four hours afterwards'.

The exact nature of declination is a subject of much research; in some models, it is treated by allowing the baseline (or both baseline and top-line) to decrease slowly over the utterance. In ToBI-like models, this downdrift in F0 is modeled by two separate components; in addition to declination, certain high tones are marked as carrying downstep. Each downstepped high accent causes the pitch range to be compressed, resulting in a lowered topline for each such accent.