Data-driven techniques

Prominence prediction by deterministic means is actually one of the most successful uses of non-statistical methods in speech synthesis. This can be attributed to a number of factors, for example the fact that the rules often don t interact or the fact that many of the rules are base on semantic features (such that even if we did use a data driven technique we would still have to come up with the semantic taxonomy by hand). Sproat notes [410] that statistical approaches have had only limited success as the issue (especially in compound noun phrases) is really one of breadth and not modelling regardless of how the prominence algorithm actually works, what it requires is a broad and exhaustive list of examples of compound nouns. Few complex generalisations are present (what machine learning algorithms are good at) and once presented with an example, the rules are not difficult to write by hand. 

Phrasing and prominence prediction can be performed by a variety of algorithms, including simple rules, sophisticated rules, data driven techniques and statistical techniques. 

All the data driven techniques described below use algorithms to learn the mapping between characters and phonemes. They use an existing lexicon as training data, and often words in this are held out as test data and so these data driven techniques are only usable in cases where there a comprehensive lexicon is available. In general words do not have a one-to-one correspondence between characters and phonemes sometimes two or more characters produce a single phoneme (e.g. gh /17 in rough) and sometimes they don t produce a phoneme at all (e.g. the e in hate... 

A number of other data driven techniques have also been used to G2P conversion. Pagel et al [343] introduced the use of decision trees for this purpose, a choice which has been adopted as a standard technique by many researchers [249], [198]. The decision tree works in the obvious way, by considering each character in turn, asking questions based on the context of the character and then outputing a single phoneme. This can be seen as a way of automatically training the context-sensitive rules described above. 

Other data driven techniques include support vector machines [114], [111], [18], transformation based learning [67], [519] and latent semantic analysis [38]. Boula de Mareiiil et al [62] describe a formal evaluation of grapheme-to-phoneme algorithms and explains the complexities involved in ensuring that accurate and meaningful comparisons can be performed. 

The statistieal approach has been somewhat of a latecomer to grapheme-to-phoneme conversion, perhaps because of the success of other data driven techniques such as pronunciation by analogy or the impression that context-sensitive rewrite rules are adequate so long as they can be automatically trained, e.g. by a decision tree. In recent years however a number of approaches have been developed which give a properly statistical approach. 

Many data driven techniques exist, including neural networks, pronunciation by analogy and decision trees. 

Black and Hunt in fact used a linear regression technique, with features such as lexical stress, numbers of syllables between the current syllable and the end of the phrase, identity of the previous labels and so on. Once learned, the system is capable of generating a basic set of target points for any input, which we then interpolated and smoothed to produce the final FO contour. Other data driven techniques such as CART have proven suitable for S3mthesizing from AM representations [292], [340],... 

While many of the AM models and deterministic acoustic models provide useful and adequate representations for intonation, the trend is clearly towards the data driven techniques described in Section 9.6. These have several advantages besides bypassing the thorny theoretical issues regarding the true nature of intonation, they have the ability to automatically analyse databases, and in doing are also inherently robust against any noise that can occur in the data, whether it be from errors in finding FO values or from other sources. 

In recent years, the notion of explicit models has been challenged by a number of data-driven techniques which learn intonation effects from data. 

Data-driven techniques have come to dominate nearly every aspect of text-to-speech in recent years. In addition to the algorithms themselves, the overall performance of a system is increasingly dominated by the quality of the databases that are used for training. In this section, we therefore examine the issues in database design, collection, labelling and use. 

Unit selection is arguably the most data-driven techniques as little or no processing is performed on the data, rather it is simply analysed, cut up and recombined in different sequences. As with other database techniques, the issue of coverage is vital, but in addition we have further issues concerning the actual recordings. 

This can be performed in a number of ways ranging from explicit models which afford a large degree of control to data driven techniques which use video clips as basic units. 

Formant synthesis was the first genuine synthesis techniqne to be developed and was the dominant technique until the early 1980s. Formant synthesis is often called synthesis by rule a term invented to make clear at the time that this was synthesis from scratch (at the time the term synthesis was more commonly nsed for the process of reconstmcting a waveform that had been parameterised for speech-coding purposes). As we shall see, most formant-synthesis techniques do in fact use rules of the traditional form, but data-driven techniques have also been used. 

While the HMM techniques described in this chapter constitute the leading approach, other data-driven techniques have been developed. AU in a sense share the same basic philosophy, namely that it is inherently desirable to use a model to generate speech since this enables compact representations and manipulation of the model parameters, and all are attempts at solving the problems of specilying the model parameters by hand. 

All algorithms are to some extent data-driven even hand-written rules use some data , either explicitly or in a mental representation wherein the developer can imagine examples and how they should be dealt with. The difference between hand-written rules and data-driven techniques lies not in whether one uses data or not, but concerns how the data are used. Most data-driven techniques have an automatic training algorithm such that they can be trained on the data without the need for human intervention. 

---