A new method for the automatic derivation of HMM topologies is presented. At first, the speech signal is segmented into acoustical units by using an Ergodic HMM. Then, a lattice structure is built from all the observed pronunciations. The lattice is thus pruned according to an information-theoretic criterion, aiming to preserve only the more characteristic event sequences. A circuit-free HMM topology is finally built after a proper state number assignment. The method naturally permits the sharing of phonetically-motivated observation densities within different HMM and states. Results for a speaker-independent recognition experiment are given. Keywords: Hidden Markov Models, Structure Inference, Information Measures, Alternative Pronunciations.
Bibliographic reference. Falaschi, Alessandro / Pucci, Massimo (1991): "Automatic derivation of HMM alternative pronunciation network topologies", In EUROSPEECH-1991, 671-674.