Avoiding explicit labels

In conclusion, labelling should be performed in line with a module s requirements. Intuitive labelling is a useM and accurate method of labelling training data for modules. Analytical labelling is often best performed by computer, and in many cases can be avoided altogetiier by the use of hidden variables in a module. Analytical labelling by expert human labellers is best avoided if at aU possible. 

Throughout the book, we have made statements to the effect that statistical text analysis outperforms rule-based methods, or that unit selection is more natural titan formant synthesis. But how do we know this In one way or anotiier, we have evaluated our systems and come to tiiese conclusions. How we go about this is tiie topic of this section. 

Once the user requirements have been stated, we can then move on to the issue of how to test whether a system meets these. Such tests are called system tests because they test the overall performance of the system. For the above examples, a test nught include measuring how many addresses drawn from a random sample are understandable by a listener. Often the results of these tests can be surprisingly poor, so another key part of a test would be to get a human to read the same addresses. We may then find that the human struggles also, and the comparison between hiunan and machine may give a much better indication of the usefulness of a TTS system in an application. 

---

Graphs—Graphs which are produced in several colors should be avoided as Notebook entries, because color distinctions vanish in the mircofilmed record. Where colors exist, the color-coded curves or areas or legends should also bear explicit labels. 

In this volume dedicated to Yngve Ohm we feel it is particularly appropriate to extend his ideas and merge them with the powerful practical and conceptual tools of Density Functional Theory (6). We extend the formalism used in the TDVP to mixed states and consider the states to be labeled by the densities of electronic space and spin coordinates. (In the treatment presented here we do not explicitly consider the nuclei but consider them to be fixed. Elsewhere we shall show that it is indeed straightforward to extend our treatment in the same way as Ohm et al. and obtain equations that avoid the Bom-Oppenheimer Approximation.) In this article we obtain a formulation of exact equations for the evolution of electronic space-spin densities, which are equivalent to the Heisenberg equation of motion for the electtons in the system. Using the observation that densities can be expressed as quadratic expansions of functions, we also obtain exact equations for Aese one-particle functions. 

Figure 9.3. Basis vectors used to describe the atomic displacements in the MLg molecule, showing the numbering system used for the seven atoms. (Labeling of the unit vectors at atoms 4, 5, and 6 is not shown explicitly to avoid overcrowding the figure.)...

---