Pitch accents

Figure 6.1 Example of FO contour with pitch accents and boundary tones maiked.

Beyond this, we again find that there is httle agreement about how to describe pitch accents and boundary tones. Some theories state there are a fixed inventory of these, while some describe them with continuous parameters. The nature of pitch accents and boundary tones is disputed, with some theories describing them as tones or levels while others say their characteristic properly is pitch movement. One prominent theory states that we have an intonational phonology that parallels normal phonology, and as such we have inventories of contrasting units (sometimes called... 

Of all the prosodic phenomena we have examined intonational tune is the most heavily related to augmentative and particularly affective content. In situations where these effects are absent, we can say to a first approximation that all utterances have in fact the same intonational tune the only differences occur as to where the pitch accents and boundary tones which make up this tune are positioned. Hence we can almost argue that for discourse neutral synthesis, there simply isn t any intonational tune prediction to be done. In other words, the real task is to predict a suitable FO contour that e q)resses the prominence and phrasing patterns and encodes the suprasegmental, rather than true prosodic patterns of the utteranee. 

In practice standard regression algorithms are difficult to use directly for FO prediction, because of a serious mismatch in the nature of the input and output representations. In general the input representation is a representation of intonational form as just discusses (for example, pitch accent types and positions) whereas the output is a continuous list of real valued numbers. In particular, the feature combination needs to generate not one, but a sequence of FO values, which is further complicated because the number of values in this sequence can vary from case to case ... 

Much of the discussion on the subject of tune centres around how to deseribe pitch accents. A pitch accent is commonly manifested in the FO contour as a (relatively) sudden excursion from the previous contour values. This is where association comes in, as pitch accents only occur in conjunction with prominent syllables, and in doing so attract attention to that syllable. Pitch accents can only occur in association with prominent syllables (see Section 6.3 on prominence), but need not occur on all prominent syllables. 

Many theories of intonation define pitch accents as the fundamental unit of intonation. While we defer specific systems of pitch accent descriptions for the section below on intonational models, here we attempt to describe some pitch accent phenomena in theory neutral terms. In most models that use piteh accents, we see a basic distinction between the pitch accent s inherent properties and its setting. 

In general, by inherent properties, we mean ffie shape of the FO contour of the pitch accent. To a degree, we can separate the FO pattern of the pitch accent from the surrounding contour, but... 

The version of the theory presented in t Hart and Cohen [446] describes contours in terms of three declination lines - high, middle and low. Pitch accents are realised by rising and falling between these declination lines. An example of a stylized and standardized contour is shown in Figure 9.9 (fi-om Willems [494]). 

Figure 9.12 A pitch accent split into rise and fall components.

Some models have a superimpositional nature (Dutch, Fujisaki) where pitch accents are seen as being relatively short term which riding on top of a phrases which are seen as being relatively long. By contrast, in linear models (AM/Tilt) contours are composed of linear sequences of in-... 

The biggest difficulty with the AM models is the difficulty in labelling corpora. Several studies have been conducted in labelling with ToBI with the general conclusion that while labellers can often identify which syllables bear pitch accents, they are very poor at agreeing on which particular... 

Two major problems stem from this. Firstly, any database which has been labelled with ToBI will have a significant amount of noise associated with the pitch accent label classes. Secondly, for any large scale machine learning or data driven approach, we need a considerable amount of labelled data to the extent that it is impractical to label data by hand. As we shall see in Chapters 15 and 16, virtually all other aspects of a modem data driven TTS system s data are labelled automatically, and so it is a significant drawback if the intonation component can not be labelled automatically as well. Because however the level of human labeller agreement is so low, it is very hard to train a system successfully on these labels we can hardly expect an automatic algorithm to perform better than a human at such a task. 

One solution that is increasingly adopted is to forgo the distinction between label types altogether, see for instance [488]. While the break index and boundary tone components are often kept, only a single type of pitch accent is used in effect the labellers are marking whether a word is intonationally prominent or not. However, it should be clear that such an approach effectively reduces ToBI to a data driven system of the type described below. 

The Fujisaki model is most commonly used with Japanese, but has been used or adapted to many other languages. In Japanese, we find that the range of pitch accent phenomena is narrower than in languages such as English, which means that the model s single type of accent is particularly suited. In addition, the nature of intonation in Japanese means that accents are marked in the lexicon, which greatly simplifies the problem of prominence prediction. Hence a simple approach to this, which uses accent information from the lexicon alone, is often sufficient. A common approach therefore is to determine phrase breaks and prominent syllables from the text, and then phrase by phrase and syllable by syllable generate the input command parameters for the Fujisaki... 

A completely comprehensive approach to timing would therefore attempt to model all these factors, but in general it is very hard to determine any reliable anchor points at a level lower than the phone. In practical TTS, the complex sub-phone timing patterns are often ignored as firstly it is difficult to predict the interactions and secondly it is subsequently difficult to know what to do with the sub-phone information once obtained. In practice then the most common approaches are to predict syllable durations or phone durations. Syllables are attractive as they seem to be the natural units of prosody we have seen elsewhere that they are the units of prominence and therefore pitch accent placement. In addition, we usually think of the syllable as being the natural unit of rhythm. Phones are attractive as they are the smallest units we can predict and therefore allow the most detail in prediction. 

The AM school (which includes ToBI) describes intonation in terms of abstract High and Low tones. Diacritics (, %, -) are used to specify which tones align with syllables and boundaries. The tones can be combined in various ways to form an inventory of pitch accents (e.g. H +L). 

One way of realising this is as a direct extension of the original diphone principle. Instead of recording and analysing one version of each diphone, we now record and analyse one version for each combination of specified features. In principle, we can keep on expanding this methodology, so that if we wish to have phrase initial, medial and final units of each diphone, or a unit for every type or variation of pitch accent, we simply design and record the data we require. 

Great care must be taken when recording prosodic databases as the recording conditions can influence the speaker to the extent that the prosody is unrepresentative of real speech. In particular it is inadvisable to instruct the speaker to speak different types of pitch accent and so on as this will inevitably lead to artiflcial exaguration of the intonation. It can even be argued that prosody recorded in a studio will nearly always be artificial as the speaker is not engaged in a normal discourse [81],... 

Hirschberg, J. Pitch accent in context Predicting intonational prominence from text. Artificial Intelligence 63 (1993), 305-340. 

Ladd, D., Mennen, 1., and Schepman, A. Phonological conditioning of peak alignment in rising pitch accents in Dutch. The Journal of the Acoustical Society of America 107 (2000), 2685. 

Marsi, E., Busser, B., Daelemans, W., Hoste, V., Reynaert, M., and Bosch, A. V. D. Combining information sources for memory-based pitch accent placement. In Proceedings of the International Conference on Speech and Language Processing 2002 (2002). 

Oliver, D., and Clark, R. Modelling Pitch Accent Types for Polish Speech Synthesis. In Proceedings of Interspeech 2005 (1995). 

VAN Santen, J. Quantitative Modeling of Pitch Accent Alignment. International Conference on Speech Prosody. Aix-en-Provence, France (2002), 107-112. 

Yuan, J., Brenier, J., and Jurafsky, D. Pitch Accent Prediction Effects of Geme and Speaker. Proceedings ofinterspeech (2005). 

---