A Multistrategy Approach to Improving Pronunciation by Analogy

Abstract

Pronunciation by analogy (PbA) is a data-driven method for relating letters to sound, with potential application to next-generation text-to-speech systems. This paper extends previous work on PbA in several directions. First, we have included “full” pattern matching between input letter string and dictionary entries, as well as including lexical stress in letter-to-phoneme conversion. Second, we have extended the method to phoneme-to-letter conversion. Third, and most important, we have experimented with multiple, different strategies for scoring the candidate pronunciations. Individual scores for each strategy are obtained on the basis of rank and either multiplied or summed to produce a final, overall score. Five strategies have been studied and results obtained from all 31 possible combinations. The two combination methods perform comparably, with the product rule only very marginally superior to the sum rule. Nonparametric statistical analysis reveals that performance improves as more strategies are included in the combination: this trend is very highly significant (p < 0:0005). Accordingly for letter-to-phoneme conversion, best results are obtained when all five strategies are combined: word accuracy is raised to 65.5% relative to 61.7% for our best previous result and 63.0% for the best-performing single strategy. These improvements are very highly significant (p ∼ 0 and p < 0:00011 respectively). Similar results were found for phoneme-to-letter and letter-to-stress conversion, although the former was an easier problem for PbA than letter-to-phoneme conversion and the latter was harder. The main sources of error for the multistrategy approach are very similar to those for the best single strategy, and mostly involve vowel letters and phonemes.

Joint-sequence models for grapheme-to-phoneme conversion

“…In this pronunciation lattice, each node represents a candidate phoneme, and each path through the lattice represents a possible pronunciation. Marchand and Damper (2000) extend and improve this approach by combining different path scoring strategies. Yvon (1996) constructs the lattice representing all potential pronunciations of a word by extracting overlapping chunks from words in the training lexicon.…”

Introducing nativization to Spanish TTS systems

“…The latest and most successful implementation of the algorithm was published by Marchand and Damper (2000), which we have reimplemented for our experiments. This system as well as the initial one, called PRONOUNCE (Dedina and Nusbaum, 1991) consists of four major components.…”

“…Each complete path through the lattice is called a "pronunciation candidate". In this work, we considered only the shortest paths through the lattice (Marchand and Damper, 2000). If there is a unique shortest path through the lattice, it is automatically chosen as the best pronunciation and the algorithm stops.…”

“…, p k } are the phonemes comprising the pronunciation candidate, with k being the pronunciation length. Marchand and Damper (2000) proposed to use five scoring strategies in order to choose the best pronunciation. Also, two methods of strategy combination were introduced.…”

Automatic acquisition of grammatical types for nouns

“…(Rentzepopoulos and Kokkinakis, 1996) describes a hidden Markov model approach for phoneme-tographeme conversion, in seven European languages evaluated on a number of corpora. (Marchand and Damper, 2000) uses a fusion of data-driven and pronunciation-by-analogy methods, obtaining word accuracies of 57.7% and 69.1% for phoneme-tographeme and grapheme-to-phoneme experiments respectively, when evaluated on a general dictionary. (Llitjos and Black, 2001) report improvements on letter-to-sound performance on names by adding language origin features, yielding 61.72% word accuracy on 56,000 names.…”