The system we are presenting here uses audio input to detect the topic of the conversation. In the current version we use a speech recognition software built with the IBM's ViaVoice SDK, which is run in a continuous dictation mode, producing a list of candidate words in the real time. This list of most recently spoken words is matched against a set of trained topic models to provide likelihoods of the sequence of words in each of the topics. Topic probabilities are compared to each other and the maximum one is concluded to represent the most likely current topic of the conversation.