SPEECH OGLE Indexing Uncertainty for Spoken Document Search.pdfVIP

SPEECH OGLE: Indexing Uncertainty for Spoken Document Search Ciprian Chelba and Alex Acero Microsoft Research Microsoft Corporation One Microsoft Way Redmond, WA 98052 chelba, alexac@ Abstract applications that range from web and private net- work search to searching for private information re- The paper presents the Position Specific siding on one’s hard-drive. Posterior Lattice (PSPL), a novel lossy Speech search has not received much attention representation of automatic speech recog- due to the fact that large collections of untranscribed nition lattices that naturally lends itself spoken material have not been available, mostly to efficient indexing and subsequent rele- due to storage constraints. As storage is becoming vance ranking of spoken documents. cheaper, the availability and usefulness of large col- In experiments performed on a collec- lections of spoken documents is limited strictly by tion of lecture recordings — MIT iCam- the lack of adequate technology to exploit them. pus data — the spoken document rank- Manually transcribing speech is expensive and ing accuracy was improved by 20% rela- sometimes outright impossible due to privacy con- tive over the commonly used baseline of cerns. This leads us to exploring an automatic ap- indexing the 1-best output from an auto- proach to searching and navigating spoken docu- matic sp