Paper: Multi-Speaker Language Modeling

ACL ID N04-4034
Title Multi-Speaker Language Modeling
Venue Human Language Technologies
Session Short Paper
Year 2004

In conventional language modeling, the words from only one speaker at a time are repre- sented, even for conversational tasks such as meetings and telephone calls. In a conversa- tional or meeting setting, however, speakers can have signi cant in uence on each other. To recover such un-modeled inter-speaker in- formation, we introduce an approach for con- versational language modeling that considers words from other speakers when predicting words from the current one. By augmenting a normal trigram context, our new multi-speaker language model (MSLM) improves on both Switchboard and ICSI Meeting Recorder cor- pora. Using an MSLM and a conditional mu- tual information based word clustering algo- rithm, we achieve a 8.9% perplexity reduction on Switchboard and a 12.2% reduction on the ICSI M...