1. Introduction

Interactive information providing systems or reservation systems that use edited recorded human voices as outputs to users have recently become more popular in a variety of fields. In implementing these kinds of systems, different types of speech synthesis methods can be used depending on the quality of service desired. Automated directory assistance systems [1] or ticket reservation systems often use dialogue templates with slots that are filled with spoken words selected from a pre-recorded database. Such systems need very large speech databases to provide sufficient coverage. Conventional interactive systems using dialogue templates suffer from the following difficulties:

1.

The narrator who recorded the original set of templates may not be available when a set of new words is to be recorded.

2.

It is almost hopeless to ask a single narrator to pronounce all words in a uniform tone and power if the number of words is extremely large, such as all city and town names in a country. Even if it is possible to do so, it would take too long to complete the set, occasionally up to several years.

3.

Traditional speech synthesis methods, like formant synthesis, can yield a large number of voices at a reasonable cost. However, the quality of the resulting voices is not acceptable in most cases.

Under the above conditions, service developers inevitably use several narrators to pronounce different parts of the database as uniformly as possible. This increases cost dramatically. The method proposed in this paper solves these problems in a cost-efficient way. It improves a conventional method by introducing a new concept: mora position together with mora length can used instead of the prosodic features adopted by conventional methods. The proposed method is tested in an automated directory assistance service in Japan; it synthesizes most Japanese city or town names and was found to yield good synthesis quality at the word level.