Impact of microphone on computer applications with voice input modality

In the paper we investigate the impact of different microphones on computer applications that employ voice input, like applications and tools developed for motor-handicapped persons. In order to make their comparison objective, test utterances were recorded always via a pair of microphones, a reference one and a test one. Using this setup, sixty half-minute-long recordings were acquired from 6 persons and via 6 different microphones. These were used in speech recognition tests that revealed quite large and surprisingly inconsistent impact of the microphones. To compensate the variability between them we examined several techniques for channel and speaker adaptation. With the best one (known as Constricted Maximum Likelihood Linear Regression) we achieved an average absolute improvement of 2.7 % in a voice dictation task.