IMPROVED MONOSYLLABIC WORD MODELING ON SWITCHBOARD EXECUTIVE SUMMARY

The SWITCHBOARD (SWB) Corpus consists of 2430 conversations digitally recorded over long distance telephone lines. The SWB Corpus totals over 240 conversation hours (elapsed time) of data. The average conversation duration is six minutes. The transcriptions contain more than 3 million words of text. The SWB Corpus includes more than 500 adult-aged speakers and covers most major American English dialects. Such impressive statistics make SWB the premier database for telephone bandwidth large vocabulary conversational speech recognition (LVCSR) research. The goal of this project is to resegment the speech data and correct the transcriptions in an effort to significantly advance LVCSR technology. We have completed the first six months of the SWB project and have released 525 conversations with corrected segmentation, transcriptions, and automatic word alignments. Additionally, there are 275 conversations awaiting release with automatic word alignments. These 800 conversations comprise 41 % of the conversations used in the WS’97 partition, and 33 % of the entire SWB corpus. We have also performed a major overhaul of the lexicon by removing incorrect or unnecessary entries and making the lexicon case sensitive. Finally, we have created extensive documentation including a statistical analysis of the conversations and a description of th