Retrieving a User Language Model from an Unsupervised Document Map
Mikko Kurimo and Krista Lagus
In: Machine Learning Meets the User Interface, 12 December 2003, Whistler, Canada.
This work presents a method to automatically retrieve a language model focused on the topic and style of the speech situation at hand. The retrieval is based on a sample text or a first-pass transcription hypothesis of the speech. We use a self-organizing map (SOM) of all the training texts to index and define the different language models by the map nodes. The smoothly organized index enables a fast local search and easy access to models of different topical granularity.