Data driven approaches

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 facts that the rules often don t interact and that many of the rules are based 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 because the issue (especially in compound-noun phrases) is really one of breadth, not modelling regardless of how the prominence algorithm actually works, what it requires is a broad and exhaustive hst 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. 

That said, with the provision of huger, labelled corpora, the natural progression has still been towards using machine-learning techniques to predict prominence. We should probably not be surprised to find that these techniques follow the basic paradigm used for 

---

We use a method that implements the Unbiased Prediction Risk criterion [13] to provide a data-driven approach for the selection of the regularization parameter. The equality constraints are handled with LQ factorization [14] and an iterative method suggested by Villalobos and Wahba [15] is used to incorporate the inequality constraints [10]. The method is well suited for the relatively large-scale problem associated with analyzing each image voxel as no user intervention is required and all the voxels can be analyzed in parallel. 

The randomisation test proposed by Wiklund et al. [34] assesses the statistical significance of each individual component that enters the model. This had been studied previously, e.g. using a t- or F-test (for instance, Wold s criterion seen above), but they are all based on unrealistic assumptions about the data, e.g. the absence of spectral noise see [34] for more advanced explanations and examples. A pragmatic data-driven approach is therefore called for and it has been studied in some detail recently [34,40]. We have included it here because it is simple, fairly intuitive and fast and it seems promising for many applications. 

The data-driven approach does not render the model-driven approach obsolete or superfluous. In fact, data-driven approaches have to rely, and will do so for the foreseeable future, on model-driven advances in single fields of science (Aebersold, 2000). In addition, without knowledge gained with reductionist approaches interpretation of data with the systems approach is not possible. Novel tools such as pathway analysis and pattern recognition, and novel kinds of research questions such as the investigation of hypotheses based on the interaction of several system components, become possible with the systems approach. 

Early approaches to fault diagnosis were often based on the so-called physical redundancy [11], i.e., the duplication of sensors, actuators, computers, and softwares to measure and/or control a variable. Typically, a voting scheme is applied to the redundant system to detect and isolate a fault. The physical redundant methods are very reliable, but they need extra equipment and extra maintenance costs. Thus, in the last years, researchers focused their attention on techniques not requiring extra equipment. These techniques can be classified into two general categories, model-free data-driven approaches and model-based approaches. 

Model-free data-driven approaches do not require a model of the monitored process, but only a good database of historical data collected in normal operating conditions. This class of approaches includes both statistical and knowledge-based methods [49],... 

Whereas hard filters can be considered to be knowledge-driven, soft filters are the result of a data-driven approach. A quantitative structure-activity or structure-property relationship (QSAR/QSPR) is established to predict a property from a set of molecular descriptors. Examples are the above-mentioned in-silico prediction tools for frequent hitters [27] and drug-likeness [41,42] additional models for ADM E properties are described below. 

The two examples above show that the structure of a problem determines what type of search to use. In the first example, there is only a limited number of components which could be broken, and a large number of possible symptoms. A data-driven approach to this problem would be to ask the scientist for all the symptoms, some of which may require disassembling the HPLC to determine. The problem, therefore, demands a goal-directed approach. 

Church, A. H., Walker, A. G., Brockner, J. (2002). Multisource feedback for organization development and change. InJ. Waclawski A. H. Church (Eds.), Organization development A data-driven approach to organizational change (pp. 27-54). San Francisco Jossey-Bass. 

Organization development A data-driven approach to organizational change (pp. 3-26). San Francisco Jossey-Bass. 

A considerable number of references discuss imaging algorithms that are entirely data driven and do not require a structural model. Similar to the passive damage diagnostic approach, NN have been proposed that would be trained on a number of damage scenarios and then used to identify an actual damage however, this data-driven approach seems to require a large number of tests and could be cost prohibitive. 

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. 

The unit selection synthesis technique described in Chapter 16 uses an entirely data driven approach, whereby recorded speech waveforms are cut up, rearranged and concatenated to say new sentences. Given the success of this approach in normal synthesis, a number of researchers have applied these algorithms to FO synthesis [296], [310], [311]. 

The differences in the second generation techniques mainly arise from how explicitly they use a parametrisation of the signal. While all use a data driven approach, some use an explicit... 

Second generation S5mthesis systems are characterised by using a data driven approach to generating the verbal content of the signal. 

Mori, H., Ohtsuka, T., and Kasuya, H. A data-driven approach to source-formant type text-to-speech system. In Proceedings of the International Conference on Speech and Language Processing 2000 (2000). 

The final further issue concerns the specification-to-parameters problem. In going from formant synthesis to LP synthesis the key move was to abandon an explicit specification-to-parameters model of vocal-tract configurations and instead measure the required parameters from data. From this we can use at synthesis time a lookup table that simply lists the parameters for each unit (typically a diphone). The cost in doing so is firstly that we lose explicit control of the phenomenon and secondly of course that we incur significant extra storage costs. If we now, however, look at how closely the LP parameters for one diphone follow parameters for lots of examples of that same diphone in many different real situations, we see that in fact even this purely data-driven approach is severely lacking, and in many cases the set of parameters for our diphone will match... 

---