Audio retrieval of internet public opinion is an important component of public opinion monitoring and management. It can improve the utilization and retrieval rate. This paper proposes a retrieval system of internet audio public opinion based on keyword spotting which consists of audio public opinion collection, the dynamic reference template definition and online search. The keyword spotting model uses the dynamic technology of keywords increment to generate dynamic vocabulary to achieve the keywords of public opinion online decision-making and flexible definition of reference templates. Last, the mixed confidence algorithm is designed to confirm the results. Experimental results show that this keyword spotting based on dynamic reference template can automatically complete the training and optimization of system via the accumulation of retrieval times. On the other hand, the mechanism also improves the system detection rate.