Indexing spoken audio by LSA and SOMs

This paper presents an indexing system for spoken audio documents. The framework is indexing and retrieval of broadcast news. The proposed indexing system applies latent semantic analysis (LSA) and self-organizing maps (SOM) to map the documents into a semantic vector space and to display the semantic structures of the document collection. The SOM is also used to enhance the indexing of the documents that are difficult to decode. Relevant index terms and suitable index weights are computed by smoothing the document vectors with other documents which are close to it in the semantic space. Experimental results are provided using the test data of the TREC's spoken document retrieval (SDR) track.