Exploring Grapheme-to-Phoneme Induction with Machine Learning

Text-to-speech (TTS) systems have increasingly found use in the modern world. One of the subproblems of TTS is determining the phonetic structure of words, i.e., their pronunciation, from their orthography, i.e., their spelling. This is known as the grapheme-to-phoneme (G2P) problem. In all languages this is a nontrivial task, but particularly in English, a language with rich historiolinguistics that has led to an irregular and inconsistent spelling system. A single letter, even appearing in similar contexts, can be pronounced several different ways. For example, see Table 1. The most common solution to the obstacle of unpredictable pronunciations is using a phonetic dictionary with entries that look like the rows in Table 1. However, neologisms like Google in English, Wikcionario in Spanish, and Klimakatastrophe in German constantly creep into the lexicon. Keeping track of all these new words is impossible, yet native speakers can easily pronounce most neologisms at first sight. This suggests that their oththography carries enough information to determine their pronunciation. The first solutions to the G2P problem involved hand-writing sets of pronunciation rules for various languages. With the onset of the machine learning (ML) paradigm, this method became extremely outmoded.[4] In the ML version, the basic idea is to develop a set of rules which map strings of graphemes to strings of phonemes based on their orthographic context, i.e., the letters surrounding the grapheme. These rules are learned from a phonetic dictionary like [2] which can then be applied to new words outside of the training set.