Research On Query-By-Example Key Word Spotting System

Key word spotting is a special speech recognition mechanism.It combines many technologies such as signal processing and pattern recognition.It is widely applied in human-computer interaction,audio information retrieval,interception and security.Query-by-Example key word spotting system is one of the important research directions in the field of key word spotting,and it uses template to retrieval keywords,thus having its own advantages in some applications.This paper mainly focuses on the methods based on text and template matching.The main contents of this paper are as follows,(1)The performance of phoneme recognizer based on LC-RC(Left and Right Temporal Contexts)system was tested,and the experimental results were analyzed.(2)A text-based key word spotting system based on Kaldi platform was built.The graph-based rescore algorithm was implemented,which used the similarity between the hypothetical intervals to resort and filter the scores of the recognition results.The performance of the algorithm was tested and proved to be improved.(3)The existing key word spotting system based on template matching was improved.Firstly,the method of threshold setting was studied,and then a template score fusion method based on Logistic regression was implemented.Finally,the method of score normalization was compared and analyzed.