Recognition-based word counting for reliable barge-in and early endpoint detection in continuous speech recognition

In this paper, we present a word counting method that enables speech recognition systems to perform reliable barge-in detection and also make a fast and accurate determination of end of speech. This is achieved by examining partial recognition hypotheses and imposing certain “word stability” criteria. Typically, a voice activity detector is used for both barge-in detection and end of speech determination. We propose augmenting the voice activity detector with this more reliable recognition-based method. Experimental results for a connected digit task show that this approach is more robust for supporting barge-in since it is less prone to interrupting the announcement when extraneous speech input is encountered. Also, by using the early endpoint decision criterion, average response times are sped up 75% for this connected digit task.