Using Intelligent Environments for Language Model Selection

About 1 in 5 Americans have some kind of disability, and 1 in 10 have a severe disability (McNiel, 1997). Intelligent environments can augment lost abilities, but they depend upon speech recognition. Unfortunately, current speech recognition systems are not accurate enough for common use. A home environment was instrumented with home automation devices, motion detectors, and a speech recognition system, which logged data for approximately half a month. Three language models were generated, based upon three command categories. We compare a reinforcement learning approach and a probabilistic approach towards using information from the motion detectors to automatically select the most relevant language model. As expected, using initial data the probabilistic approach returned better results, with an error rate of 47.24%. The reinforcement learner had similar results, with an error rate of 53.14%. We look forward to reporting future results using more reasonable amounts of data.