Position Specific Posterior Lattices for Indexing Speech

Proceedings of the 43rd Annual Meeting of the ACL, pages 443–450, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Position Specific Posterior Lattices for Indexing Speech Ciprian Chelba and Alex Acero Microsoft Research Microsoft Corporation One Microsoft Way Redmond, WA 98052 {chelba, alexac}@microsoft.com Abstract The paper presents the Position Specific Posterior Lattice, a novel representation of automatic speech recognition lattices that naturally lends itself to efficient in- dexing of position information and subse- quent relevance ranking of spoken docu- ments using proximity. In experiments performed on a collec- tion of lecture recordings — MIT iCam- pus data — the spoken document rank- ing accuracy was improved by 20% rela- tive over the commonly used baseline of indexing the 1-best output from an auto- matic speech recognizer. The Mean Aver- age Precision (MAP) increased from 0.53 when using 1-best output to 0.62 when us- ing the new lattice representation. The ref- erence used for evaluation is the output of a standard retrieval engine working on the manual transcription of the speech collec- tion. Albeit lossy, the PSPL lattice is also much more compact than the ASR 3-gram lat- tice from which it is computed — which translates in reduced inverted index size as well — at virtually no degradation in word-error-rate performance. Since new paths are introduced in the lattice, the OR- ACLE accuracy increases over the origi- nal ASR lattice. 1 Introduction Ever increasing computing power and connectivity bandwidth together with falling storage costs re- sult in an overwhelming amount of data of vari- ous types being produced, exchanged, and stored. Consequently, search has emerged as a key applica- tion as more and more data is being saved (Church, 2003). Text search in particular is the most active area, with applications that range from web and in- tranet search to searching for private information re- siding on one’s hard-drive. Speech search has not received much attention due to the fact that large collections of untranscribed spoken material have not been available, mostly due to storage constraints. As storage is becoming cheaper, the availability and usefulness of large col- lections of spoken documents is limited strictly by the lack of adequate technology to exploit them. Manually transcribing speech is expensive and sometimes outright impossible due to privacy con- cerns. This leads us to exploring an automatic ap- proach to searching and navigating spoken docu- ment collections. Our current work aims at extending the standard keyword search paradigm from text documents to spoken documents. In order to deal with limitations of current automatic speech recognition (ASR) tech- nology we propose an approach that uses recogni- tion lattices — which are considerably more accu- rate than the ASR 1-best output. A novel contribution is the use of a representation of ASR lattices which retains only position informa- tion for each word. The Position Specific Posterior 443 Lattice (PSPL) is a lossy but compact representa- tion of a speech recognition lattice that lends itself to the standard inverted indexing done in text search …

Content truncated to 500 words