Lattice-based statistical spoken document retrieval

Abstract

Recent research efforts on spoken document retrieval (SDR) have tried to overcome the low quality of 1-best automatic speech recognition transcripts -- especially for conversational speech -- by using statistics derived from speech lattices containing multiple transcription hypotheses as output by a speech recognizer. However, these efforts have invariably used the classical vector space retrieval model. In this thesis, I present a lattice-based SDR method based on a statistical approach to information retrieval. I formulate a way to estimate statistical models for documents from expected word counts derived from lattices; query-document relevance is computed as a log probability under such models. Experiments show that my method outperforms statistical retrieval using 1-best transcripts, a recent lattice-based vector space method, and BM25 using lattice statistics. I also extend my proposed SDR method to the task of query-by-example SDR -- retrieving documents from a speech corpus, where the queries are themselves full-fledged spoken documents (query exemplars).