Automatic Speech Recognition and Instant Translation using Multi Agent System and Data Mining

Abstract

Speech recognition is a process of converting signal of speech into a sequence newlineof word. Neural Network and dynamic programming and there are several various newlinemethod used for the process of speech recognition. The speech recognition system is newlineto recognize the word and responds to user with visual feedback based on the newlineaccuracy level of the user s speech. The discrete modules of the speaker identification newlinesystem for speech processing are sample training, speaker database and classification. newlineThis system, training the input speech sample collections from 8 no of various newlinespeaker. In the signal processing model, a set of feature vectors has produced from the newlineapplied signal for every speaker specifically whereas mathematical model get fitted newlinewith feature vector set. Using vector quantization code book feature vectors are newlineextracted and get mapped on to HBM state. Once it is obtained as the feature vectors, newlinewe built the acoustic model representation. The statistical distribution of vector newlinequantization (VQ) model is used for every speaker features. The codebook gets newlinesubstituted to every vector set features that has smallest code vector sets with fixed newlinesize. Codebook get collected in the database of speaker but the design of codebook newlinehas general goal in order to reduce the training data in quantization distortion. newlineTherefore, code vector get substituted to their adjacent neighbors have generated in newlinethe codebook. During recognition mode signal processing technique is used as newlinetraining for the process of input speech sample whereas the feature get quantified in newlinethe database using of each codebook. Hence, the codebook provides least distortion to newlinethe speaker which is acknowledged. Today speech recognition system is based on the newlinemodel of Ti-HBM. It has the capability to model the duration of acoustic unit namely newlineprobability of survival which it is used as the built in parameter. These parameters has newlineacquired from the joint state-time of parameter distribution. Ti-HBM parameters are newlinetrained by using Viterbi algorithm has been proposed. newlineIn this research, we have reduced the training duration span by implementing newlinethe Time-inhomogeneous hidden Bernoulli model (TI-HBM) to overcome Hidden newlineMarkov Model (HMM) for speech recognition. TI-HBM is basically a Bernoulli newlineprocess where it can able to model easier and quicker than HMM in acoustic-unit newline(phonemes/word) duration by using an inbuilt framework named survival probability newlinev newlinefor speech recognition. To evaluate the recognized word by using the comparison of newlineHBM and HHM based on the performance metrics such as Word Error Rate (WER), newlineWord accuracy, Word efficiency and time period. In terms of word recognition newlineaccuracy, the TI-HBM outperforms the HMM. Here all the simulations are done by newlineMATLAB. newline