Automatic Speech Recognition for Mobile Communication Devices

The implementation and evaluation of an automatic speech recognition (ASR) based task for a mobile, hand-held device is presented. A description of the task is provided along with a set of compensation techniques that are used to compensate speaker independent hidden Markov models (HMMs) with respect to environment and transducer variability. A technique for combined environment/transducer compensation is presented for this task and is shown to significantly reduce the effects of environmental mismatch. The overall performance degradation was reduced from 43.2 percent to 13.5 percent for speech spoken through a far--field microphone in an office environment, and from 81.2 percent to 45.8 percent for the same transducer in a noisy cafeteria environment.