Phonetisaurus

Why we included this project

Phonetisaurus generates pronunciation dictionaries for speech recognition by training grapheme-to-phoneme models on OpenFst. It provides C++ binaries with Python bindings, so you can train models and then use them to produce pronunciations for words your lexicon lacks, which saves the manual work of hand-curating entries. When alignments need closer inspection, it can also dump the raw per-word lattices in .fst format. This is a focused library rather than a turnkey package: plan on building OpenFst yourself and knowing your way around a Kaldi-style pipeline. Its long history in the speech community is the main reason teams still choose it despite the setup effort.