Agent/Client Speech Identification for Mixed-Channel Conversation in Customer Service Call Centers
Van Hai, Van Tuan Mai · 2020
In our recent work, we proposed a method to use automatic speech recognition (ASR) to monitor customer service call centers. Specifically, text decoded by ASR is used as the input for a natural language understanding unit to achieve phone call categories, phone call scores or detect abnormal phone calls, etc. To do this, we need to have two separated audio streams from call centers i.e., one for agent, one for client side. However, in many call centers in Vietnam, the agent and client audio streams are mixed together due to many reasons such as hardware capacity, storage saving. In this paper, we propose a method to identify agent and client speech in a mixed-channel conversation. We first cluster speech segments in a mixedchannel conversation into two clusters using acoustic information. After that, language modelling technique works together with an ASR system to identify the agent/client label for each segment in the conversation. Experimental results show that the proposed method works effectively with over 95% of accuracy for both agent and client sides. Our approach has been applied successfully in several call centers in Vietnam.