7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Compact Subnetwork-Based Large Vocabulary Continuous Speech Recognition

Dong-Hoon Ahn, Minhwa Chung

Sogang University, Korea

We present an improved method of compactly organizing the decoding network for a semi-dynamic network decoder. In the previous work [1], the network management units called subnetworks were made compact by self-structuring themselves. We improve this subnetwork representation in two aspects by employing the shared-tail topology [2]. Firstly, we localize the decoding algorithm so that it works with a set of subnetworks rather than with the whole decoding network. Secondly, we align unshared suffixes of pronunciations into a shared tail to reduce redundancies. Experimental results on a 20k-word Korean dictation task show that our algorithm significantly reduces the memory requirement and produces additional gains in word accuracy by using aligned shared tails.

Full Paper

Bibliographic reference.  Ahn, Dong-Hoon / Chung, Minhwa (2002): "Compact subnetwork-based large vocabulary continuous speech recognition", In ICSLP-2002, 725-728.