DESIGN OF SPEAKER INDEPENDENT TURKISH SPEECH CONTROL SYSTEM IN NOISY VEHICLE ENVIRONMENT

摘要

This paper represents the results of speaker independentrnTurkish speech control experiments in vehiclernenvironments. Almost there are unlimited numbers ofrnwords in the Turkish language, it is impossible to usernwords as the basic unit in the system. In this case,rnassuming to use the sub-units is more efficient, we choosernto work with context-dependent (different vocabularyrncontext in training and test) tri-phones as the smallestrnunits and modelled with Hidden Markov Model (HMM).rnAlso to limit the complexity of the tri-phone modelsrndecision tree based state clustering is used. Proposedrnspeech recognition system is able to recognize the speechrnwaveform by translating the speech waveform into a setrnof feature vectors using Mel Frequency CepstralrnCoefficients (MFCC) that are the typical recognitionrnparameters. In experiments on hands-free in-car speechrnrecognition with the microphone far from the talker, thisrnframework is found to be effective in terms of recognitionrnrate and computational cost under various driving speeds.rnTo examine the recognition performance of the system,rntri-phone based acoustic model is tested with differentrndecision tree pruning factors. System experiments resultsrnhad shown that the word correctness of system tests isrnbetween 50-86 percent. This ratio is fulfills the safetycriticalrnoperations requirements of a vehicle.