The architecture of a system for full-text search by speech data based on a global search index

Abstract

This paper presents the architecture of a system for full-text search by speech data based on a global search index that combines information about all speech recordings in the archive. The architecture includes two independent blocks: an indexing block, and a block for building and performing a search query. In order to process speech recordings, it uses an automatic speech recognition system (ASR) with a linguistic decoder based on weighted finite-state transducers framework (WFST), which generates word lattices. Lattices are sequentially converted to confusion networks and inverse indexes. It allows taking into account all the word hypotheses generated during decoding. The proposed solution expands the applicability of speech analytics systems for those cases when the word error rate is high, such as the processing of speech recordings collected under difficult acoustic conditions or in low-resource languages.