Improving letter-to-pronunciation accuracy with automatic morphologically-based stress prediction

Robust text-to-speech (TTS) systems require a letter-topronunciation module for generating the pronunciations of words missing from the system lexicon. These pronunciations must specify not only the phone sequence that corresponds to an input orthography, but also the location of lexical stress. However, letter-to-pronunciation modules that make use of a window of context letters around a target letter normally cannot “see” larger-context morphological information that is highly correlated with stress location. The present work demonstrates that by adding a new component that uses morphological information to predict which letter of a word might receive primary stress, and then using the resulting “stressed letters” as input to a decision tree stressed-letter-topronunciation component, improvements to both stress accuracy and phone accuracy are obtained in American English, British English, and German. Furthermore, using stressed letters as the input to the decision tree also improves phone accuracy when stress is not required in the output pronunciation, as is conventionally the case for automatic speech recognition (ASR).