Research And Implementation Of Keyword Spotting System With Large Keyword Table In Spontaneous Speech

As an important way to human-computer interaction,speech recognition has broad application prospects.Compared with the continuous speech recognition, keyword spotting technology has advantages of high correct rate and strong practicality, so keyword spotting has become a research hotspot in speech recognition.In this dissertation,we implement a keyword spotting system in spontaneous speech, which has features of speaker independent,task independent and large keyword table.We firstly introduce some relevant technologies used in keyword spotting system, including speech pre-processing and feature parameter extraction,HMM model, language model,keyword search and keyword verification.Then,according to our goal, we design the system structure,including pre-processing and feature parameter extraction,speech recognizer,keyword search and confidence measure.In the module of keyword search,we discuss two kinds of speech recognition results:N-Best lattice and Syllable lattice,and the experiment shows the keyword search based on syllable lattice is better than N-Best lattice in detection rate and speed.In the module of confidence measure,we describe three common confidences: acoustic confidence,confidence based on post-probability and confidence based on dynamic ranking information,and analyze their characters by experiments.On this basis, two kinds of mixture confidences are presented,they are mixture confidence based on acoustic confidence and post-probability,and mixture confidence based on dynamic ranking and post-probability.Experiments show that both of mixture confidences have better performance than single confidence,and the mixture confidence based on acoustic confidence and post-probability has a better performance than another mixture confidence.On the basis of researches above,we implement a keyword spotting system adopting the best method shown by our experiments.In a task of 500 keywords, system's FOM reaches 74.50%in telephone-channel spontaneous speech。...