Syllable-based Language Models in Speech Recognition for English Spoken Document Retrieval

The spoken content of audio/visual collections such as TV or radio archives is an information resource of enormous potential. The challenge is to develop methods that will make it possible to browse or search these collections. The experimental results presented in this paper demonstrate that syllable-level transcripts provide an important supplement to conventional word-level transcripts for the task of unlimited vocabulary American English spoken document retrieval. Recognition is performed with syllable language models with vocabulary sizes 20k, 10k, 5k, 1k, and 500. The syllable recognition rates of the 10k and 5k models are comparable to that achieved by a baseline 100k word-based language model. A simple retrieval experiment involving a fuzzy full text search supplies proof-of-concept that syllable-based transcripts make it possible to retrieve spoken documents that contain query words not included in the 100k vocabulary of the word-based language model.