음성인식 신뢰도를 위한 BMS알고리즘 연구

In this paper, we proposed BMS(Background Model Set) algorithm working speech recognition to compensate calculating shortcoming of conventional existent RLJ-CM(RLJ-Confidence Measure) and NCM(Normalized Confidence Measure). Confidence Measure displays relative likelihood between recognized models and unrecognized models. CM brought bad result in calculating probability and standard deviation using all phoneme at the process of building anti-phone model. Also, there was shortcoming that recognition time increases at the calculation using all phoneme. In order to solve this problem, we studied about method to reconstitute average and standard deviation taking BMS algorithm. Using BMS algorithm, FAR is 0.104 FA/KW/HR (false alarm/keyword/hour) in MDR(Missed Detection Rate) 17% neighborhood. This experiment improved 50% than used before BMS algorithm. Also, we improved 33% recognition time that the average time of recognition was 10 minutes using BMS than 15 minutes at the process database for speaker 1 person's estimation in isolated word recognition.