CVC-based network representation of phoneme sequences with applications to spoken word recognition

This paper describes a procedure to incorporate a rule-based network representation into automatic speech recognition systems. The network is used for representing acoustic-phonetic variations of spoken words. Rules for generating the networks are formulated in the framework of CVC-type units, and derived from a fine-labeled speech database. An automatic word recognition system is constructed to test the effectiveness of this method. The recognition results indicate that test samples are rarely mis-recognized when they are adequately labeled into acoustic-phonetic segment sequences and the network representation is substantially useful to satisfy this condition.