Speech retrieval using spoken keyword extraction and semantic verification

Speech is an informative signal, which contains not only the meaning of its text content but also the emotions of the speaker represented by prosodic information. This study presents a novel approach using spoken keyword extraction and semantic verification to speech retrieval. An information extraction approach based on acoustic, prosodic and linguistic information is proposed for significant information/keyword extraction. Semantic verification with forward-backward propagation is applied to improve the precision of the retrieved documents. In the forward procedure, a bag of keywords is selected according to word significance measure. In the backward procedure, semantic relation between keywords is estimated for semantic verification. The verification score is then used to weight and re-rank the retrieved documents to improve the retrieval performance. Experiments were conducted on a collected broadcast news database of 198 hours. Experimental results show that the proposed methods achieve better retrieval results compared to other conventional indexing methods.