Barge-in robust spoken dialogue interface using multichannel sound field control and array signal processing

A spoken dialogue system is demanded as a user-friendly human-machine interface that does not require any special skills in its manipulation. Speech has advantageous features: they are hands-free and eyes-free, i.e., one can use speech while doing other tasks. For effective utilization of the features, it is desirable that the system can be used even when the user stands away from the microphone or the user’s speech is uttered interrupting the output sound of the system (response sound). The problem in satisfying such demands is the degradation of automatic speech recognition (ASR) because of feedback of response sound and observation of interfering noise due to other sound than the user’s speech. Since current ASR systems are sensitive to noise, a noise reduction method is indispensable. In elimination of the response sound and the interfering noise, an acoustic echo canceller (AEC) and an adaptive beamformer (ABF) are generally used, respectively. In each of the methods, a filter is adapted to eliminate its target noise based on the minimum-mean-squared-error criterion. Thus, when their filters are trained using signals containing sources other than their target noise, their performances degrade severely. To prevent such degradation, the system should detect the times when the observed signals contain sounds other than the target noise, denoted as double-talk detection (DTD). However, accurate DTD is difficult, particularly in such a situation that both response sound and interfering ∗Doctoral Dissertation, Department of Information Processing, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0561031, September 30, 2007.