AdolfVonKleist/Phonetisaurus

Implements grapheme-to-phoneme conversion using weighted finite-state transducers (WFSTs) built on OpenFst, enabling training and inference of pronunciation models from lexicons. Provides C++ binaries for model training and compilation alongside Python 3 bindings for programmatic access to alignment scores, lattice extraction, and n-best candidate generation. Integrates with speech recognition frameworks like Kaldi and supports flexible output modes including probability mass constraints and normalized probability scoring.

510 stars. No commits in the last 6 months.