Effects of acoustic transmission characteristics upon word recognition performance

The effects of various acoustic transmission characteristics on speech recognition accuracy are described. There are two main factors that determine the acoustic quality of a speech recognition system: the presence of acoustically reflective surfaces and the frequency response of the microphone. The effects of dip level and dip frequency, caused by a reflective surface, on a DP-based word recognition method are measured using an omnidirectional microphone. For perfectly reflective surfaces it is shown that word recognition can be degraded by up to 20% by reflected wave interference. The frequency responses of four kinds of directional microphone are measured in various acoustic conditions. The measurements indicate that there are three types of frequency response. Incorrect microphone usage yields poor frequency response and significantly degrades word recognition accuracy. To overcome the problems caused by reflected waves, two techniques are proposed: the power weighted likelihood ratio distance measure and multiple templates. As confirmed by tests, the former can improve recognition accuracy by 10% while the latter offers a 20% improvement.<>