Speech recognition technology can help searching and understanding the contents of audio-visual records that archives hold. But old video records sometimes do not guarantee good recognition results due to low signal quality or lack of vocabulary used at that time. This paper shows actual experimental results and trials to enhance the accuracy using speech recognition toolkit based on deep learning, by training with relevant corpus data for video records in the 1950s and 1970s. This paper also proposes a strategy for records management applications, considering of accuracies and service purposes for the future.