8.2.2 Names

As the error analysis above indicated, names are an important issue in speech synthesis. The many types can be categorized into personal names (first names and surnames), geographical names (city, street, and other place names), and commercial names (company and product names). For personal names alone, Spiegel (2003) gives an estimate from Donnelly and other household lists of about two million different surnames and 100,000 first names just for the United States. Two million is a very large number; an order of magnitude more than the entire size of the CMU dictionary. For this reason, most large-scale TTS systems include a large name pronunciation dictionary. As we saw in Fig. 8.6 the CMU dictionary itself contains a wide variety of names; in particular it includes the pronunciations of the most frequent 50,000 surnames from an old Bell Lab estimate of US personal name frequency, as well as 6,000 first names.

How many names are sufficient? Liberman and Church (1992) found that a dictionary of 50,000 names covered 70% of the name tokens in 44 million words of AP newswire. Interestingly, many of the remaining names (up to 97.43% of the tokens in their corpus) could be accounted for by simple modifications of these 50,000 names. For example, some name pronunciations can be created by adding simple stress-neutral suffixes like s or ville to names in the 50,000, producing new names as follows:

walters = walter+s lucasville = lucas+ville abelson = abel+son

Other pronunciations might be created by rhyme analogy. If we have the pronunciation for the name Trotsky, but not the name Plotsky, we can replace the initial /tr/ from Trotsky with initial /pl/ to derive a pronunciation for Plotsky.

Techniques such as this, including morphological decomposition, analogical for-

mation, and mapping unseen names to spelling variants already in the dictionary (Fackrell and Skut, 2004), have achieved some success in name pronunciation. In general, however, name pronunciation is still difficult. Many modern systems deal with unknown names via the grapheme-to-phoneme methods described in the next section, often by building two predictive systems, one for names and one for non-names. Spiegel (2003, 2002) summarizes many more issues in proper name pronunciation.