Improving ASR Performance with OCR Through Using Word Frequency Difference

Recently, there has been a growing interest in conversational artificial intelligence (AI). As a result, research is actively being conducted on automatic speech recognition (ASR) to facilitate interactions between humans and machines. This paper proposes a system that enhances ASR performance. The proposed method accumulates images captured from lecture videos in real-time every 30 seconds. The frequency ratios between text data from captured images and text data calculated offline from over 333K are used to improve the ASR performance. Experimental results showed that the word error rate (WER) decreased by a maximum of 0.68% compared to using only the traditional ASR. Especially, the recognition rate for specialized terms frequently used in lectures showed an improvement of 64%.