Research And Application Of Key Word Spotting Technology

Key-word Spotting has been a hot research topic in speech recognition field, which has broad application prospects. In recent years,Key-word Spotting technologies based on template matching have been the mainstream approaches in this field, these approaches locate the occurrences of a spoken key-word by matching against the key-word audio template and test audio utterance directly based on the acoustic similarity matching. Different from the traditional Key-word Spotting method based on the LVCSR system, such Key-word Spotting technologies avoid the training data problem and the out-of-vocabulary(OOV) problem, and have good extensibility. The key considerations in those Key-word Spotting technologies based on template matching are template representation and template matching algorithm, the research of template representation is concentrated on posteriorgram features, Dynamic Time Warping (DTW) template matching techniques and its variants have been the mainstream matching methods, these two key aspects will be analyzed and studied in this paper.This paper focus on the study and analysis of the main Key-word Spotting technologies based on template matching algorithms, with having carried on the system implementation and experimental analysis.The main work in this paper includes:1. This paper studies and implements the Key-word Spotting algorithm based on the Gaussian posteriorgram feature and the Segmental-DTW matching algorithm, and has carried out the system performance experiments based on a variety of speech corpus.2. Phonetic posteriorgram feature extraction procedures based on Lattice are mainly analyzed in this paper, and the Key-word Spotting algorithm based on the phonetic posteriorgram are preliminarily designed and implemented. Testing experiments with comparison analysis are carried out on both English and Chinese corpus.3. We proposed the modified Key-word Spotting system, we use modified Gaussian posteriorgram based on the proposed Gaussian components selection algorithm as template representation, the selection algorithm is inspired by the TF-IDF concept. The key-word verification part based on the long-term prosodic feature is introduced to further improve the system performance, and we proposed the Fisher vector encoding method for 3-dimensional symbol prosodic feature. With the experimental verification, the proposed key-word Spotting system make a certain performance improvement.