Re-Engineering Letter-to-Sound Rules

Re-Engineering Letter-to-Sound Rules Martin Jansche The Ohio State University Columbus, OH 43210, U.S.A. jansche.1@osu.edu Abstract Using finite-state automata for the text analysis component in a text-to-speech system is problem- atic in several respects: the rewrite rules from which the automata are compiled are difficult to write and maintain, and the resulting automata can become very large and therefore inefficient. Converting the knowledge represented explicitly in rewrite rules into a more efficient format is difficult. We take an indirect route, learning an efficient decision tree rep- resentation from data and tapping information con- tained in existing rewrite rules, which increases per- formance compared to learning exclusively from a pronunciation lexicon. 1 Introduction Text-to-speech (TTS) systems, like any other piece of sophisticated software, suffer from the shortcom- ings of the traditional software development pro- cess. Highly skilled developers are a costly re- source, the complexity and sheer size of the code involved are difficult to manage. A paradigmatic example of this is the letter-to-sound component within the text analysis module of a mature large- scale text-to-speech system. In the system described in (Sproat, 1998) text analysis is performed using finite-state transducers compiled from rewrite rules (Kaplan and Kay, 1994; Mohri and Sproat, 1996) and other high-level descriptions. While the exclu- sive use of finite-state technology has advantages, it is not without its shortcomings, both technical and stemming from the use of hand-crafted rule sets and how they are represented: 1. Extensive rule sets need to be constructed by human experts, which is labor-intensive and expensive (Sproat et al., 1998). 2. Realistic rule sets are difficult to maintain be- cause of complex interactions between serially composed rules. 3. Although rewrite rules can, in principle, be compiled into a huge monolithic transducer that is then very time-efficient, in practice this is not feasible because of the enormous sizes of the resulting machines (cf. the numbers given in (Mohri and Sproat, 1996) and (Sproat et al., 1998, 74)). 4. For reasons of space efficiency, certain com- putations are deferred until run-time (Mohri et al., 1996; Mohri et al., 2000), with a significant impact on time efficiency. While there is a clear need for human expert knowledge (Sproat et al., 1998, 75ff.), those experts should not have to deal with the performance as- pects of the knowledge representation. Ideally we would like to use a knowledge representation that is both time and space efficient and can be con- structed automatically from individually meaning- ful features supplied by human experts. For practi- cal reasons we have to be content with methods that address the efficiency issues and can make use of explicitly represented knowledge from legacy sys- tems, so that moving to a new way of building TTS systems does not entail starting over from scratch. As a case study of how this transition might be achieved we took the letter-to-phoneme rules for French in the TTS system described in (Sproat, 1998) and proceeded to 1. Construct a lexicon using the existing system. 2. Produce an alignment …

Content truncated to 500 words