Fast Vocabulary-Independent Audio Search Based on Syllable Confusion Network Indexing in Mandarin Spontaneous Speech

This paper presents a fast vocabulary-independent audio search method in Mandarin spontaneous speech which is based on syllable confusion network (SCN) indexing. Confusion network is linear and naturally suitable for indexing. The feasibility of using syllable confusion network as lattice representation is firstly investigated. Since direct syllabic decoding may not have a very high accuracy, long- span decoding units such as syllable collocation and word are also explored respectively. To handle spontaneous speech with high error rate and irregular prosody, syllable confusion matrix (SCM) is applied here to calculate the relevance score. Experiments carried out on conversational corpora for the keyword spotting task in the Chinese 2005 863 Evaluation show that this method can not only yield highly compact SCN lattices with syllable graph density (SGD) of 3.83, but also achieve an equal error rate (EER) of 32.45%, which is about 33% relatively reduction when comparing with the baseline top-1-based search method with an EER of 48.72%.