Added missing memory reservations for interpolated lookahead models. Also added pseudo-lookahead for htk lattice grammars (needed to keep the token count sensible in large lattices). The lattices could be handled more elegantly by actually building the grammar network in the decoder, but this improves the current quick-and-dirty solution.
Added missing memory reservations for interpolated lookahead models. Also added pseudo-lookahead for htk lattice grammars (needed to keep the token count sensible in large lattices). The lattices could be handled more elegantly by actually building the grammar network in the decoder, but this improves the current quick-and-dirty solution.