Research And Implementation On Chinese Speech Keywords Spotting Based On HMM

The speech is the most convenient means for human to communicate with each other.So making the machine understand human language or accordance with the operation of the human is a long-standing dream.Keywords spotting is a special branch of speech recognition its main task is detecting the specific terms from the continue voice.For it's consuming less resources,high-accuracy,high-practical,key word spotting has bright future in many application areas.Hidden Markov Model (HMM) is a popular model of the semantic recognition system and best tool to analysis the time-varying quasi-stationary signal.This article systematically introduced the principles and framework of keywords spotting,the theory of HMM, speech feature extraction,and fractal dimension algorithm.On this base,this paper designed a keywords spotting system on the base of HMM and make a lot of improving to the training samples requirements,filler model and detection efficiency which are major problems of the existing system.(1) Basic theory of HMM is described,and deeply analysis the former-after algorithm,the best path search algorithm of Viterbi,the multi-output Gaussian mixture model parameters re-evaluation algorithm with multi observation sequence based on the Baum-Welch.these do some basis work for the speech recognition.(2) The analysis method based dynamic frame-length was proposed.First, estimated duration of the current speech units,then determined window length in accordance with the duration.To some extent,this program alleviated the deterioration of the performance of the keywords spotting system which because of the quality and the number of training samples.And it could achieve the adaptive of speech speed when it used in detection process.(3) The structure and types of filler model are researched.And learned the advantages and disadvantages of final clustering model and syllable lattice clustering model,the new filler model program based on syllable clustering was proposed, which is well to improve system performance. (4) The first syllable model of keywords and keywords confirming model are Constructed.First,use filer model and the first syllable model of keywords to match input voice clips,if the best matching results fall into keywords domain,while the two current syllable witch maybe the candidates for the keywords match with the confirming model of this kind of keywords and calculate the average frame likelihood scores,statistics in this article more than 8 is the passing score.(5) An improved algorithm for the calculation of fractal dimension is proposed after demonstrating the principle of speech fractal dimension,and according with it,I had realize real-time syllable segmentation.Then,an online Chinese-mark and two-state decode algorithm were designed.(6) A small amount of training samples and testing samples library are set up, Establish and implement a keyword detection system based HMM without grammar restrictions,and do a comprehensive analysis of systemic.